Explore AI Models

GPT-6.1 Sol is an upgrade to GPT-6 Sol, positioned below flagship GPT-6 Astra in the GPT-6 series, released at DevDay 2026. OpenAI says it nearly matches Astra on agentic coding, computer use, and professional work at one-fifth Astra's standard token prices, while keeping GPT-6 Sol's input/output rates unchanged and halving the cached-input price. It introduces an Ultrafast service tier delivering up to 6x faster inference in the API (up to 8x in Codex, up to 300 tokens/sec) at roughly 6x the standard price. On DeepSWE v1.1, OpenAI reports it approaches Astra-level long-running software engineering performance; on Terminal-Bench Science 0.1 it averages $5.47 per task at maximum effort.

Claude Sonnet 5.5 offers the best combination of speed and intelligence in the Claude lineup, with a 1M-token context window and fast comparative latency. It accepts text and images and returns text, with adaptive thinking on by default at high effort — the lowest setting, between_tools, turns off up-front thinking rather than disabling reasoning entirely, unlike Sonnet 5 where thinking could be fully turned off. Five breaking changes apply versus Sonnet 5: forced tool use now returns an error, thinking blocks are tied to the model and conversation, the older computer_20251124 tool is no longer accepted, the advisor tool rejects older models as advisors, and text between tool calls now arrives in thinking blocks unless a display value or between_tools is set.

Claude Opus 5.5 is the first model in Anthropic's Claude 5.5 family, performing at the level of Claude Fable 5.1 on most work while costing roughly 40% less to run than Claude Opus 5 on typical workloads, with output generated over 30% faster. It supports text and image input with a 1M-token context window, adaptive thinking that is always on (default effort: medium, and cannot be disabled as an anti-distillation measure), and incorporates Fable 5.1's safeguards for cybersecurity, biology, and distillation, detailed in a 230-page system card. Recommended as a daily driver for coding, financial and document workflows, and computer use.

GPT-6 Luna is OpenAI's fast, low-cost entry-tier model in the GPT-6 family, trained with the same methods as flagship GPT-6 Astra but tuned to run faster and cheaper. It cuts API prices 50% versus GPT-5.6 Luna's promotional pricing, with cached input tokens now discounted 90%. At higher reasoning-effort levels it matches GPT-5.6 Sol's factuality on OpenAI's internal evaluation at about a hundredth of the cost, and reaches 66.6% on DeepSWE. Free and Go ChatGPT users can access Luna through the desktop app.

GPT-6 Sol is OpenAI's mid-tier model in the GPT-6 family, sitting below flagship GPT-6 Astra. Trained using the same methods as Astra but optimized to run faster and cost less, it cuts API prices 50% versus GPT-5.6 Sol's promotional pricing, with about half the factuality errors of its predecessor and cached input tokens now discounted 90%. OpenAI reports it matches Claude Fable 5.1 xhigh on FrontierCode coding benchmarks at substantially lower cost, and scores 33.2% on AutomationBench (xhigh) versus 26.9% for Claude Opus 5 at roughly a ninth of the per-task cost.

Grok 4.7 is xAI's flagship multimodal reasoning model, built on a new, larger base model than Grok 4.6 (reportedly ~2.1T parameters vs. 1.5T) with a longer reinforcement-learning run weighted toward multi-hour tasks. It accepts text and images, supports a 500K-token context window, four reasoning levels (low, medium, high, xhigh), and native understanding of the Grok Bot harness for agentic coding and knowledge work. xAI reports stronger self-verification, long-context handling, and document/presentation generation, at the same per-token price as Grok 4.6.

Jev is TypeSafe's flagship System One model, served by POST /v1/systemone. It evaluates a text `state` against a map of typed `questions` (noul yes/no, choice, and score) and returns one calibrated `answer` per question. `jev-latest` currently resolves to `jev-1.13.0`. Text only (string, JSON object, or array of text values); no image, audio, or video input. Trained with RLCD for calibrated decisions; not fine-tuned on customer data. Charged per input token, output tokens are free.

GLM-5.3-FlashX is Z.ai's high-speed serving variant of GLM-5.3-Flash, delivering inference speeds of up to 200 tokens/s for faster responses. It inherits Flash's 320B-total/18B-active MoE design, native multimodality, hybrid sparse and linear attention architecture, and 1M-token context window. FlashX runs on a purpose-built serving stack (SGLang-based inference engine, W8A8 quantization, mixed INT8/FP8/BF16 cache quantization, and encode-prefill-decode disaggregation), reporting roughly 3x end-to-end throughput over Z.ai's initial baseline on the same hardware. It is suited for interactive coding agents, browser/computer-use loops, visual coding, and other multi-step workflows where latency compounds.

DeepSeek-V4.1-Flash is a multimodal Mixture-of-Experts model with a 552B-parameter backbone, activating just 8B parameters per token during prefill and 16B during decoding. It natively processes images and text, supports a 1M-token context window, and introduces a Causal Encoder-Decoder architecture with Compressed Sparse Attention 2 to make long-context inference more efficient. Configurable reasoning-effort levels let developers trade latency and cost for deeper deliberation. Compared with DeepSeek-V4-Flash-0731, it reduces the global KV-cache footprint by roughly 4x while delivering stronger coding, reasoning, and agentic performance.

GPT-Live 1 is OpenAI's full-duplex voice model for natural, expressive real-time conversations with smooth interruption handling. It listens and speaks at the same time, and delegates reasoning and tool use to a backend agent (OpenAI Responses delegation or client-managed delegation). Voice sessions are billed by duration per second; backend model and tool usage is billed separately. Accessed over WebSocket via /v1/live/sessions.

GPT Image 2.5 Sunburst is OpenAI's high-fidelity text-to-image generation and editing model, offering an extra level of precision for detailed creative work at the cost of longer generation times than Flare. It shares Images 2.5's improvements in reference fidelity, precise edits, and multi-turn consistency, with sharper detail, more natural lighting, richer textures, and better handling of complex layouts and transparent backgrounds. Recommended for premium visual workflows like production-ready campaign creative or polished product imagery.

GPT Image 2.5 Flare is OpenAI's fast text-to-image generation and editing model, delivering higher-quality images than GPT Image 2 at up to 50% lower latency. It is the default choice for most API applications, sharing Images 2.5's improvements in reference fidelity, precise edits, and multi-turn consistency, and is well suited to creator and social content, product experiences, visual search, rapid prototyping, and high-volume generation.

GPT-6 Astra is OpenAI’s flagship frontier model for complex end-to-end work, featuring very long context (up to ~1.05M tokens), advanced reasoning with configurable effort levels, and strong performance on coding, computer use, research, and professional tasks. It supports text and image inputs with text-only outputs, tool use via the Responses API (including web search, code execution, file search, and computer use), and structured outputs. Astra is heavily safety-aligned and designed for multi-step workflows, async tool calling, and mid-turn steering in demanding enterprise and developer applications.

Gemini 3.8 Flash is Google’s most capable Flash-series multimodal model, optimized for long-horizon software engineering, autonomous agents, and complex enterprise workflows while maintaining high speed and cost efficiency. It supports very long contexts (up to 1,048,576 tokens) across text, images, video, audio, and PDFs, and offers advanced capabilities like tool calling, computer use, file search, and structured outputs. It is suitable for agentic applications, complex reasoning, and high-throughput production workloads.

Claude Fable 5.1 is Anthropic's generally available, production-safeguarded deployment for demanding reasoning and long-horizon agentic work, sharing underlying weights with Claude Mythos 5.1. It extends Claude Fable 5 with stronger long-running agentic coding, multistep research, and document, spreadsheet, and slide work, plus per-message effort control, turn-scoped system messages, readable in-progress tool-call updates, and content provenance. Adaptive thinking is always on, with the effort parameter controlling reasoning depth (default: high).

GLM-5.3-Flash is Z.ai’s flagship natively multimodal GLM-5-series model with a 1M-token context window, accepting text plus images and video and producing text outputs. It targets fast, high-quality general and coding assistance with strong reasoning at “flash” latency and is released as open weights under the MIT license.

GLM-5.3 is Z.ai’s latest GLM-5-series flagship, a 743B-parameter text-only foundation model post-trained from GLM-5.2 for frontier-level coding, long-horizon agentic workflows, and defensive cybersecurity analysis. It inherits GLM-5’s 200K-token context window and up to 128K-token outputs, supports advanced reasoning and tool use, and is initially exposed through Z.ai’s coding plan and forthcoming API access.

Gemini 3.7 Flash is a highly capable, natively multimodal Gemini 3-series model optimized for fast, agentic workflows and complex reasoning across text, images, video, audio, and PDFs. It supports large 1M-token contexts, advanced tools like function calling, file search, computer use, and structured outputs, and configurable "thinking" levels for more deliberate reasoning. Designed for production applications via the Gemini API, it balances speed, cost-efficiency, and strong multimodal understanding.

Grok 4.6 is xAI's frontier multimodal reasoning model for coding, long-running agentic tasks, knowledge work, and interactive/visual projects, released as the successor to Grok 4.5. It accepts text and images and returns text, with a 500,000-token context window, a February 1, 2026 knowledge cutoff, and four reasoning-effort levels (low, medium, high, xhigh, with high as default). It supports function calling, JSON-schema structured outputs, web search, X search, code execution, and prompt caching, and received a longer supplemental training run than 4.5 with curated technical and agentic reasoning data.

Qwen-Image 3.0 Pro is the high-quality tier of the Qwen-Image 3.0 series, offering stronger text rendering, more realistic textures, and better semantic adherence for both text-to-image (T2I) and image-to-image/editing (I2I). Output resolution ranges from 512x512 up to 2048x2048 (PNG).

Seedance 2.5 is ByteDance's flagship multimodal video generation model, producing up to 30-second clips with native synchronized audio in a single pass. It accepts text, image, and reference-video/audio inputs, supports up to 30 image references, 10 video references, and 10 audio references, and offers real-person reference and timestamp-level editing/extension, aimed at cinematic storytelling and reference-driven creation.

DeepSeek V4 Flash is an efficiency-optimized Mixture-of-Experts large language model from DeepSeek with a 1M-token context window and strong general, coding, and reasoning performance. It is designed as a fast, low-cost frontier model suitable for long-context chat, agents, and tools. The Flash variant emphasizes high throughput and low latency while maintaining competitive reasoning quality.

MiniMax-H3 is an omni-modal video generation model that produces 4-15 second clips at up to 2K resolution with native synchronized stereo audio in a single pass. It accepts a text prompt plus optional first-frame and last-frame images, and up to 9 reference images, 3 reference videos, and 3 reference audio clips, making it suited to storyboard continuation, style and subject transfer, and audio-driven performance.

GPT Transcribe is OpenAI's speech-to-text model for completed audio files, streamed file transcripts, and committed turns in Realtime sessions over WebSocket. It supports unstructured context, keyword hints, and multiple language hints to improve transcription of domain terms, multilingual audio, and code-switching, and reports which languages it actually detected. On OpenAI's Common Voice benchmark across 22 languages, it cuts word error rate to 19.27% against 40.37% for whisper-1. It returns plain text only, without subtitle formats, word/segment timestamps, or speaker labels.

Claude Opus 5 is Anthropic’s latest flagship reasoning model in the Claude Opus family, offering near-frontier intelligence with strong performance on complex analysis, coding, and multi-step reasoning tasks. It supports very long-context interactions (up to around 1M tokens) and is designed as a general-purpose assistant for enterprise and developer use via the Claude API. Pricing matches the prior Opus 4.8 generation, with significantly higher quality per dollar than earlier Opus models.

Gemini 3.6 Flash is a multimodal frontier-intelligence model optimized for speed and cost, supporting text, image, video, audio, and PDF inputs with text-only outputs. It targets complex agentic workflows, strong code generation, and spatial/multimodal reasoning while maintaining a 1M-token context window and large output capacity. The model supports advanced tools such as function calling, computer use, file search, search grounding, structured outputs, and thinking.

Gemini 3.5 Flash-Lite is a low-latency, cost-effective multimodal model optimized for high-throughput, low-cost execution in agentic workflows and document parsing. It supports text, image, video, audio, and PDF inputs with text-only outputs, a 1M-token context window, and up to 65,536 output tokens. The model is designed for subagent tasks, simple data extraction, and applications where latency and API cost are primary constraints.

Qwen-Image 3.0 is a general-purpose image generation and editing model that balances quality and speed. It excels at complex text rendering (multi-line, paragraph-level layouts), fine detail, and realistic textures, and supports both text-to-image (T2I) and image-to-image/editing (I2I) on DashScope's multimodal-generation endpoint.

Grok Imagine Video 1.5 is xAI’s next-generation AI video model for animating still images into short, cinematic clips. It preserves the source image’s composition and identity while adding motion, camera movement, and synchronized audio such as dialogue, effects, ambience, and music.

Kimi K3 is Moonshot AI’s 2.8-trillion-parameter flagship multimodal reasoning model with a 1M-token context window, designed for long-horizon coding, complex knowledge work, visual reasoning, and tool-using workflows. It uses Kimi Delta Attention and Attention Residuals to support efficient million-token contexts and always runs in high-effort reasoning mode. The model is exposed via an OpenAI-compatible chat completions API.

Gemma 4 26B is a 31B-parameter dense multimodal model from Google DeepMind, handling text, image, and audio input to generate text output. It supports an explicit thinking mode via control tokens, tool use, and a 256K-token context window, and is released under an Apache 2.0 license.

NVIDIA Nemotron 3 Nano 30B is a hybrid Mamba-2/MoE/Attention model with 3.5B active parameters, trained from scratch as a unified model for both reasoning and non-reasoning tasks. Reasoning traces can be toggled via a chat template flag, and the model supports tool use within a 1M-token context window, making it well suited for specialized agentic applications.

NVIDIA Nemotron 3 Super is a 120B-parameter open mixture-of-experts model that activates just 12B parameters per token, delivering high compute efficiency and strong reasoning, agentic, and conversational capabilities for complex multi-agent applications. It supports tool use, thinking/reasoning traces, and a large 262K-token context window, and is optimized for collaborative agents and high-volume workloads such as IT ticket automation.

GPT-5.6 Luna is OpenAI’s fast, cost-efficient tier in the GPT‑5.6 family, optimized for low-latency, high-volume workloads such as drafting, summarization, and routine automation. It supports large-context multimodal chat and tool use with a reported 1.5M-token context window and standard OpenAI function-calling/tooling semantics. Luna trades some peak capability for significantly lower cost compared to Sol and Terra, making it suitable as a default production workhorse model.

gpt-oss-120b is OpenAI's larger open-weight model, designed for powerful reasoning, agentic tasks, and versatile developer use cases. It offers native function calling, python tool calls, structured outputs, full chain-of-thought access, and configurable reasoning effort (low/medium/high), released under the Apache 2.0 license. The MoE weights are post-trained with MXFP4 quantization for efficient deployment.

gpt-oss-20b is OpenAI's smaller open-weight model, designed for lower latency, local, or specialized use cases while retaining strong reasoning and agentic capabilities. It supports native function calling, python tool calls, structured outputs, full chain-of-thought access, and configurable reasoning effort, released under the Apache 2.0 license.

GPT-5.6 Terra is the balanced, mid‑tier model in OpenAI’s GPT‑5.6 family, designed for everyday professional and enterprise workloads. It supports text and image inputs with text outputs, offers strong reasoning and coding capabilities comparable to GPT‑5.5 at roughly half the cost, and is available in a limited preview via the OpenAI API. Terra is intended as a cost‑efficient default for broad knowledge work, analysis, and software development tasks.

GPT-5.6 Sol is OpenAI’s flagship frontier model in the GPT‑5.6 family, offering advanced reasoning, coding, and cybersecurity capabilities with multimodal text and image input and text output. It is released in a limited preview for trusted partners via the OpenAI API, with enhanced safety systems for higher‑risk and sensitive use cases. The model is designed for complex professional work, long‑horizon tasks, and agentic workflows, with improved robustness and prompt caching over prior generations.

Grok 4.5 is xAI’s flagship multimodal reasoning model optimized for coding, agentic workflows, and knowledge work, with support for very long contexts (up to ~500K tokens) and tools. It accepts text and images and can call tools for search, code execution, and other actions via the xAI API. Designed as an Opus-class model, it targets high performance at aggressive per-token pricing for large, tool-heavy sessions.

GPT-Realtime-2.1 is OpenAIs realtime reasoning model for complex voice-agent workflows.

Gemini 3.1 Flash-Lite Image is a cost-efficient multimodal Gemini 3.1 variant optimized for high-volume image workflows, supporting text and image inputs with both image and text outputs. It is suitable for image generation, editing, and related multimodal tasks where low latency and price are important. The model is available via the Gemini API in Google AI Studio and for enterprises via Vertex AI.

Claude Sonnet 5 is Anthropic’s latest mid-tier, highly agentic Claude model with a 1M-token context window, designed for everyday work, coding, tool use, and long-running agents. It supports text and image inputs and can generate long, structured text outputs with improved reliability and safety over Sonnet 4.6. It is optimized for cost-effective large-context reasoning and workflow automation across chat, API, and enterprise integrations.

Gemini Omni Flash is a preview model designed for fast, conversational video generation and editing. It turns text, images, video, and audio into 720p video (3–10 seconds, 24 FPS) and supports iterative refinement through the Interactions API. Input modalities include text, image, video (up to 10s for editing), and audio. Tasks: text_to_video, image_to_video, reference_to_video, and edit. Aspect ratios: 16:9 (default) and 9:16. Delivery: base64 or uri. No free tier.

GLM 5.2 is a large-scale reasoning model from Z.ai. It supports text input and output with a 1M-token context window, and is suited for long-horizon agent workflows, project-level software engineering,...

Kimi K2.7 Code is a coding-focused model in Moonshot AI's Kimi K2 family, built for reliable end-to-end programming tasks over long contexts. It supports text and image input, always operates in thinking mode, and is designed for long-horizon coding, agentic task decomposition, tool use, and multi-turn workflows.

Anthropic/Claude Fable 5 is Anthropic’s newest generally available, public-facing Mythos-class model, positioned as its most capable model for complex, long-running work. According to Anthropic and reporting on the launch, it is built for software engineering, knowledge work, vision, research, analysis, and long-horizon autonomy, with strong safety guardrails that block or route some high-risk requests to Claude Opus 4.8.

MiniMax M3 is a multimodal foundation model from MiniMax, built for coding, agentic workflows, long-context reasoning, and multimodal inputs. It supports up to 1 million tokens of context through MiniMax’s Sparse Attention (MSA) architecture and is positioned as a frontier model for long-horizon tasks.

Claude Opus 4.8 is Anthropic's most capable generally available Opus-family model, built for complex reasoning, long-horizon agentic coding, advanced analysis, knowledge work, vision understanding, and memory-heavy workflows. It supports text, image, and file inputs with text output, includes reasoning support, and offers a 1M-token context window for handling very large prompts, documents, and multi-step tasks.

grok-build-0.1 is xAI's specialized, ultra-fast AI model explicitly engineered for agentic software engineering and multi-step coding workflows. Rather than operating as a general-purpose conversational chatbot, it is optimized to run tools, handle debugging, and execute complex programming instructions autonomously.

Gemini 3.5 Flash is Google's high-efficiency multimodal model optimized for speed, coding, and agentic workflows. It supports text, image, video, audio, and PDF inputs with text output, and provides strong reasoning/tool-use at lower latency.

x-ai/grok-imagine-video is xAI’s video generation model for creating short clips from text or images, with native synchronized audio such as dialogue, sound effects, and music. It supports text-to-video, image-to-video, and video editing/extending workflows, making it useful for cinematic clip generation and iterative creative editing.

Grok Imagine Image is xAI’s text-to-image generation model for creating visuals from prompts, with support for both text and image inputs. xAI describes it as a model for generating images through the Grok Imagine API, and its docs list it as available for image generation with configurable output sizes.

Grok-4.3 is xAI's advanced reasoning model, launched in late April 2026, designed for complex tasks like advanced logic, math, scientific analysis, and multi-step agentic workflows.

OpenAI's GPT-5.5 Pro is a premium variant of the GPT-5.5 model, designed for higher-tier users (Pro, Business, Enterprise) handling demanding workloads like advanced research, business analysis, legal tasks, and deep reasoning with superior accuracy and depth. It builds on GPT-5.5's agentic capabilities but adds extended reasoning modes for complex, multi-step workflows.

OpenAI's GPT-5.5 is a large language model (LLM) released on April 23, 2026, codenamed "Spud," positioned as the company's smartest model yet for complex real-world tasks like coding, research, data analysis, and agentic workflows. It features variants including GPT-5.5 Thinking and GPT-5.5 Pro (not available to free-tier users), with API access starting April 24 after safeguards implementation

DeepSeek-V4-Pro is a 1.6 trillion parameter Mixture-of-Experts (MoE) language model from DeepSeek AI, with 49 billion parameters activated per token and support for a 1 million token context window. It features a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), which reduces inference FLOPs to 27% and KV cache to 10% of DeepSeek-V3.2 at 1M context, enabling 83.5% MRCR comprehension. The model was pre-trained on over 32 trillion tokens using the Muon optimizer and a two-stage post-training process involving SFT, RL with GRPO, and on-policy distillation for domain-specific expertise.

OpenAI's gpt-image-2 (also referred to as GPT Image 2 or ChatGPT Images 2.0) is a state-of-the-art text-to-image generation and editing model that transforms natural-language prompts into high-quality, photorealistic visuals with exceptional prompt fidelity, accurate text rendering, and advanced reasoning capabilities.

Moonshot AI's Kimi-K2.6 is a frontier-scale, open-source Mixture-of-Experts (MoE) multimodal model with 1 trillion total parameters (32B active), a 256K-262K token context window, and native support for text, image, and video inputs, excelling in long-horizon coding, agentic workflows, multi-agent orchestration, and complex software engineering.

Opus 4.7 is Anthropic’s next-generation Opus model, designed for long-running, asynchronous agents. Building on the coding and agentic capabilities of Opus 4.6, it offers improved performance across complex tasks.

zai-org/GLM-5.1 is Z.AI's (formerly Zhipu AI) open-weight, instruction-tuned coding flagship model—a refreshed upgrade over GLM-5—excelling in agentic engineering, complex system programming, and long-horizon tasks with SOTA open-source performance approaching Claude Opus 4.6.

Gemma 4 26B A4B IT is a 26B-parameter Mixture-of-Experts, instruction-tuned, open-weight multimodal model from Google DeepMind that processes text, image, and video inputs to produce text outputs. It targets long-context (up to ~256K tokens) reasoning, coding, and chat use cases while remaining efficient via 4B active parameters. The model is suitable for agents, RAG, and tool-calling style applications where high quality and large context are required.

google/gemma-4-31b-it is Google DeepMind’s instruction-tuned 31B Gemma 4 model, built for strong reasoning, coding, agentic workflows, and multimodal understanding. It supports text and image input, has a 256K context window, and is positioned as the dense, higher-quality counterpart to the 26B MoE variant.

MiniMax-M2.7 is a next-generation large language model designed for autonomous, real-world productivity with self-improvement capabilities. Created on March 18, 2026, it integrates advanced agentic capabilities through multi-agent collaboration, enabling it to plan, execute, and refine complex tasks across dynamic environments.

MiniMax-M2.7-HighSpeed is the ultra-fast variant of MiniMax's M2.7 flagship model, delivering approximately 100 tokens per second—3x faster than competitors like Claude Opus 4.6 (~33 tps) and GPT-5 (~40 tps)—while maintaining identical performance on complex tasks at a fraction of the cost.

GPT-5.4 nano is OpenAI's most lightweight and cost-efficient model in the GPT-5.4 family, optimized for speed-critical, high-volume tasks like classification, data extraction, ranking, and sub-agent execution. It prioritizes low latency and efficiency over deep reasoning, making it ideal for real-time systems, background tasks, distributed agent architectures, coding assistants, and multimodal applications involving images. The model supports text and image inputs (no audio or video) with text outputs, a 400,000-token context window, and up to 128,000 max output tokens; its knowledge cutoff is August 31, 2025.

GPT-5.4 mini is OpenAI's compact, high-performance model in the GPT-5.4 family, balancing advanced reasoning, multimodal capabilities, and efficiency for high-volume workloads like coding, agent workflows, and production-scale applications. It delivers significant improvements over GPT-5 mini in coding, reasoning, multimodal understanding, and tool use, running over 2x faster while approaching GPT-5.4 performance on benchmarks like SWE-Bench Pro and OSWorld-Verified. The model supports text and image inputs (no audio/video) with text outputs, a 400,000-token context window, up to 128,000 max output tokens, and an August 31, 2025 knowledge cutoff.

Grok 4.20 Beta is xAI’s newest flagship large language model that uses four specialized “agents” working together to answer complex questions more accurately and with deeper reasoning than prior Grok versions.

GPT-5.4 Pro is OpenAI's most advanced model, building on GPT-5.4's unified architecture with enhanced reasoning capabilities for complex, high-stakes tasks. It features a 1M+ token context window (922K input, 128K output) with support for text and image inputs. Optimized for step-by-step reasoning, instruction following, and accuracy, GPT-5.4 Pro excels at agentic coding, long-context workflows, and multi-step problem solving.

GPT-5.4 is the newest flagship model from OpenAI, merging the Codex and GPT product lines into one unified system. It supports a context window exceeding 1 million tokens — with up to 922K tokens for input and 128K for output — and accepts both text and image inputs. This allows it to handle high-context reasoning, code generation, and multimodal analysis all within a single workflow. The model shows notable improvements in areas such as coding, document comprehension, tool integration, and instruction adherence. It is built to serve as a reliable default for both general-purpose and software engineering tasks, capable of producing production-ready code, aggregating insights from diverse sources, and carrying out intricate multi-step processes with reduced iterations and improved token efficiency.

Google's gemini-3.1-flash-image-preview (codenamed Nano Banana 2) is a preview version of the Gemini 3.1 Flash Image model, optimized for high-speed image generation and editing with Pro-level quality, balancing performance, low latency, and cost efficiency.

GPT-5.3-Codex is OpenAI's most capable agentic coding model to date, combining frontier coding performance with general work automation capabilities—enabling developers and professionals to automate complex tasks across the entire software lifecycle, from code writing to infrastructure management and cybersecurity.

GPT Realtime 1.5 is OpenAI's flagship low-latency, speech-to-speech model optimized for live conversational systems, voice agents, and customer support, using persistent streaming sessions via the Realtime API (WebRTC/WebSocket).

Google's gemini-3.1-pro-preview (also called Gemini 3.1 Pro Preview) is a preview-stage, multimodal AI model from Google DeepMind, optimized for advanced reasoning, agentic workflows, software engineering, and complex problem-solving across text, images, video, audio, and PDFs. It refines the Gemini 3 Pro series with improved thinking, token efficiency, factual consistency, and reliability for multi-step tasks, tool use, and long-horizon stability. Key specs include a 1M input token context window (1,048,576 tokens), 64K-65K output tokens, and support for capabilities like function calling, structured outputs, code execution, search grounding, and a new "MEDIUM" thinking level for balancing cost, speed, and performance.

Sarvam 105B is Sarvam AI's flagship open-source Mixture-of-Experts (MoE) large language model (LLM), trained from scratch as India's first competitive 105-billion-parameter model. It delivers state-of-the-art performance across Indian languages and excels in enterprise-grade applications, with strong capabilities in multi-step reasoning, mathematics, coding, knowledge retrieval, and instruction-following.

Sarvam 105B is Sarvam AI's flagship open-source Mixture-of-Experts (MoE) large language model, trained from scratch as India's first competitive 105-billion-parameter model. It delivers strong performance across Indian languages and excels in enterprise-grade applications, with capabilities in multi-step reasoning, mathematics, coding, knowledge retrieval, and instruction-following.

Sonnet 4.6 is Anthropic’s strongest Sonnet-tier model to date, delivering near-frontier performance for coding, agentic workflows, and professional tasks. It’s especially good at iterative build cycles, navigating large codebases, managing end-to-end projects with memory, producing polished documents, and reliably using computers for web QA and workflow automation.

Seedream 5.0 Lite is ByteDance's fast, cost-efficient image-generation model from the Seedream family, built on a unified generation-editing architecture that runs text-to-image, single-image editing, multi-image fusion, and batch sequence generation through one endpoint. It supports up to 6 images per batch, reference-image fusion (image_urls), and configurable resolution up to 3K, trading some fidelity versus Seedream 5.0 Pro for faster, cheaper generation.

MiniMaxAI/MiniMax-M2.5 is a 229B-parameter Mixture-of-Experts (MoE) language model with 10B active parameters per token, designed as the world's first production-level model natively optimized for agentic scenarios like coding, tool use, search, and office productivity.

MiniMax-M2.5-HighSpeed (also called the Lightning variant) is the high-throughput version of MiniMax's 229B-parameter Mixture-of-Experts (MoE) model, delivering 100 tokens/second natively—roughly 2x faster than other frontier models—while maintaining identical capabilities to the standard 50 tokens/second version.

bytedance/seedance-2 is ByteDance’s next-generation multimodal AI video generation model for creating cinematic videos from text, images, video, and audio. It focuses on high motion quality, strong prompt adherence, realistic physics, and improved visual consistency, making it well-suited for professional video creation, storyboarding, and controllable multimedia generation.

Seedance 2 Mini (often branded as Dreamina Seedance 2.0 Mini on BytePlus ModelArk) is a low-cost, fast text-to-video and image-to-video generation model that produces short 480p or 720p clips with optional native audio. It is designed for rapid iteration and everyday production workflows where speed and cost efficiency matter more than maximum resolution or fidelity.

GLM-5 is Z.ai’s flagship open-source foundation model, built for complex system design and long-horizon agent workflows. Aimed at expert developers, it delivers production-grade results on large-scale programming tasks and competes with top closed-source models. With strong agentic planning, deep backend reasoning, and iterative self-correction, GLM-5 goes beyond code generation to help design, build, and execute complete systems end-to-end.

Saaras V3 is Sarvam AI's next-generation automatic speech recognition (ASR) model engineered for Indian languages and real-world speech conditions. Saaras V3 supports 23 languages (22 official Indian languages plus English) within a unified multilingual model. The model achieves approximately 19% Word Error Rate (WER) on the IndicVoices benchmark, improving significantly from its predecessor Saaras V2.5 which had ~22% WER. On the 10 most popular languages in the IndicVoices dataset, it achieves 19.3% WER, and performance advantages widen for lower-resource Indian languages

bytedance/seedance-2-fast is the fast-generation variant of ByteDance’s Seedance 2 video model. It’s designed to produce high-quality AI videos more quickly than the standard Seedance 2 model, making it a good fit for rapid iteration, prompt testing, and scalable content workflows.

Kling V3 is Kuaishou's latest AI video generation model, delivering high-quality, physics-accurate video synthesis with improved motion consistency, lighting realism, and creative control. It supports text-to-video and image-to-video workflows, generating clips up to 10 seconds at resolutions up to 1080p, ideal for cinematic storytelling, marketing content, and rapid prototyping.

Opus 4.6 is Anthropic’s new top-end Claude model, optimized for complex reasoning, coding, and high‑stakes professional work. It excels at agentic workflows, breaking down multi-step tasks, calling tools, and running for longer with fewer errors and less hand-holding. The model introduces a 1M‑token context window (in beta), stronger long-context retrieval, and up to 128k output tokens for large code and document workloads. It also adds “adaptive thinking” and adjustable effort levels, letting it dynamically choose when to think deeply versus respond quickly.

MoonshotAI's Kimi-K2.5 is a cutting-edge open-source multimodal large language model series from MoonshotAI, featuring a massive Mixture-of-Experts (MoE) architecture with 1 trillion total parameters—yet only 32 billion activated per token for highly efficient inference and low computational costs. It boasts 61 transformer layers, 64 attention heads, a 256K context window enabled by advanced MLA attention mechanisms, and SwiGLU activations, paired with a MoonViT vision encoder for native handling of images, videos, and cross-modal tasks like visual reasoning, UI mockup-to-code generation (e.g., React from screenshots), and agentic workflows.

GLM-Image is Z.AI's text-to-image generation model that quickly and accurately understands text descriptions to produce precise, personalized, high-quality images. It supports an 'hd' mode for more detailed and consistent output (~20s generation time) and a 'standard' mode optimized for faster generation (~5-10s), along with flexible custom resolutions from 1024px to 2048px per side.

fastrouter/auto is an enhanced meta-model that intelligently routes each prompt to the best-fit model from dozens of options — optimizing for quality, contextual relevance, and cost. You will be billed at the exact rate of the selected model, with full transparency in the Activity page and response metadata. With improved query understanding and a next-gen selection engine, Auto delivers sharper routing decisions and more consistent outputs — making it the most reliable way to get started.

Qwen Image 2512 is an improved version of Qwen Image with better text rendering, finer natural textures, and more realistic human generation. This 20B MMDiT model achieved top ranking among open-source models after 10,000 blind comparison rounds on AI Arena, released December 31, 2025, and is licensed under Apache 2.0.

Seedance 1.5 Pro is an advanced, joint audio-video AI generation model from ByteDance Seed, designed to create high-fidelity, cinematic, 4–12 second clips with synchronized audio, dialogue, and sound effects in a single pass. Utilizing a dual-branch diffusion transformer (DB-DiT), it specializes in realistic human motion, emotional nuance, and precise lip-sync for storytelling, advertising, and short-form content.

Google Veo 3.1 is a preview model code in the Gemini API for Google's Veo 3.1, a state-of-the-art cinematic video generation engine optimized for professional-grade 4K output, natively synchronized audio, and complex camera movements with high temporal consistency.

Google Veo 3.1 Fast is a generally available (GA) version of Google's Veo 3.1 Fast video generation model in Vertex AI and Gemini API, optimized for faster inference while delivering high-quality videos with synchronized native audio from text or image prompts.

Google Veo 3.1 Lite, a high-efficiency, cost-effective video generation model in the Gemini API and Vertex AI, designed for developers building high-volume applications at under 50% the cost of Veo 3.1 Fast with equivalent speed.

GLM-4.7 is Z.AI’s latest flagship model, featuring upgrades in two key areas: enhanced programming capabilities and more stable multi-step reasoning/execution. It demonstrates significant improvements in executing complex agent tasks while delivering more natural conversational experiences and superior front-end aesthetics.

Google's gemini-3-flash-preview is a high-speed, cost-effective AI model from the Gemini 3 series, designed for agentic workflows, multi-turn chats, and coding tasks. It delivers near-Pro-level reasoning and tool use with lower latency than larger variants, supporting a 1M token context window and multimodal inputs like text, images, audio, video, and PDFs.

OpenAI’s gpt-image-1.5 is a next‑generation image generation and editing model designed to create production‑quality visuals from text prompts and to edit existing images with fine control.

Wan 2.6 is a state-of-the-art, open-source multimodal AI video generation model developed by Alibaba Cloud (released around December 2025). It is designed to create high-fidelity, 1080p cinematic videos up to 15 seconds long from text, images, or reference videos.

GPT-5.2 Pro is OpenAI's flagship enterprise-grade reasoning model released on December 10, 2025, designed for sophisticated professional workflows requiring advanced problem-solving capabilities.

GPT-5.2 is OpenAI’s flagship frontier model in the GPT‑5 series, designed for the most demanding reasoning, coding, and long‑context workloads. It emphasizes stronger general intelligence, more reliable instruction following, and advanced safety compared with earlier 5.1 releases.

Seedream 4.5 is ByteDance Seed’s advanced AI image generation model for creating high-quality visuals with strong prompt adherence, excellent character consistency, and precise image editing. It supports text-to-image, image-to-image, reference-based generation, and multi-image workflows, making it useful for posters, product shots, brand visuals, illustrations, and cinematic-style imagery. It’s especially known for clear typography and readable small text, with output up to 4K resolution and strong preservation of details from reference images.

DeepSeek-V3.2 is positioned as a next-generation “general-purpose + reasoning” model intended to be a daily driver at roughly frontier (GPT‑5-class) performance on broad tasks like coding, math, agents, and general chat. It is released in several variants (such as V3.2, V3.2-Exp, and V3.2-Speciale), sharing the same core architecture but targeting different trade‑offs between efficiency and maximum reasoning power.

Kling Video O1 is the world's first unified multimodal AI video generation model developed by Kuaishou, offering an integrated platform for multiple video creation and editing tasks.

FLUX.2 Pro (also styled as Flux-2-Pro or FLUX.2 [pro]) is a production-grade AI image generation and editing model developed by Black Forest Labs, released in November 2025 as the flagship of the FLUX.2 family. It excels in creating high-quality, photorealistic images up to 4 megapixels from text prompts, with advanced controls for editing, multi-reference image use (up to 8 references), and precise outputs suitable for commercial workflows like product photography, UI prototyping, and brand design.

anthropic/claude-opus-4.5 is Anthropic’s latest frontier reasoning model, optimized for complex software engineering, agentic workflows, and long-running computer-use tasks such as multi-step automation and tool use. It is positioned as the most capable model in the Claude Opus family, offering state-of-the-art performance on challenging coding and enterprise benchmarks while being significantly more efficient than earlier Opus versions.

Gemini 3 Pro Image Preview is a high-end image generation and editing model in the Gemini 3 family (also known as Nano Banana Pro), built to produce studio-quality visuals with strong reasoning over complex prompts. It is optimized for detailed, multi-step creative workflows where you need both visual fidelity and precise control over content, layout, and text inside images.

GPT-5.1, released in November 2025, is the latest flagship large language model from OpenAI, featuring significant upgrades in intelligence, reasoning, and user experience compared to its predecessor, GPT-5.

Claude Haiku 4.5 is Anthropic’s October 2025 release of its newest compact AI language model, designed to offer near-premium intelligence, ultra-fast responses, and low operational cost. Positioned as the lightweight sibling to Claude Sonnet 4.5 and Opus 4.1, Haiku 4.5 enables “near‑frontier‑level” performance suitable for enterprise-scale workloads, free-tier operations, and real-time applications.

OpenAI's gpt-image-1-mini is a natively multimodal image model optimized for affordability and efficiency, making it suitable for projects that need high-throughput AI image generation at a lower cost. This model is designed to handle both text and image inputs and can generate new images, perform targeted image editing (like inpainting), and use reference images for style or content guidance.

Google Gemini 2.5 Flash Image (often nicknamed "nano-banana") is Google's latest state-of-the-art multimodal model for image generation and editing, designed for both developers and enterprises. It enables users to create, blend, and edit images with natural language prompts and multi-image input, all while benefiting from Gemini's deep real-world understanding and conversational editing features.

A cost-efficient version of GPT Realtime, capable of responding to audio and text inputs in realtime over WebRTC, WebSocket, or SIP connections. Supports a 32,000 token context window and up to 4,096 output tokens per response (knowledge cutoff: Oct 01, 2023). Also supports image input (text/audio/image → text/audio).

Sora-2 Pro is OpenAI’s premium offering for its latest AI-powered video and audio generation system, Sora 2, launched in October 2025. Sora-2 Pro is designed for professional creators and advanced users, giving them enhanced creative controls, access to longer and higher-quality generative video clips, and advanced editing tools.

GLM-4.6 is the latest iteration in the GLM series by Zhipu AI, designed as a large language model with about 355 billion parameters in a Mixture of Experts (MoE) architecture. It is optimized for various complex tasks including real-world coding, long-context processing, advanced reasoning, intelligent agent applications, and refined writing. The model features an expanded context window of 200,000 tokens (up from 128,000 in GLM-4.5), allowing it to handle longer and more complex interactions such as extensive documents or multi-turn conversations.

Sora-2 is OpenAI's state-of-the-art AI video and audio generation model, released in September 2025, and marks a major leap in controllable, realistic, and physically accurate video synthesis. It transforms text prompts and images into cinematic HD videos with synchronized audio, realistic motion, and advanced world simulation—enabling users to generate scenes with perfect lip-sync, sound effects, and genuine continuity.

Claude Sonnet 4.5 is Anthropic’s most advanced Sonnet model to date, optimized for real-world agents and coding workflows. It delivers state-of-the-art performance on coding benchmarks such as SWE-bench Verified, with improvements across system design, code security, and specification adherence. The model is designed for extended autonomous operation, maintaining task continuity across sessions and providing fact-based progress tracking. Sonnet 4.5 also introduces stronger agentic capabilities, including improved tool orchestration, speculative parallel execution, and more efficient context and memory management. With enhanced context tracking and awareness of token usage across tool calls, it is particularly well-suited for multi-context and long-running workflows. Use cases span software engineering, cybersecurity, financial analysis, research agents, and other domains requiring sustained reasoning and tool use.

Seedream 4.0 is a next-generation AI image generation and editing model developed by ByteDance Seed, unifying text-to-image synthesis, image editing, and multi-image composition in a single architecture for high-resolution outputs up to 4K. It excels in emotional intelligence, atmospheric rendering, character consistency, and fast inference (as quick as 1.8 seconds per 2K image), supporting resolutions like 2048×2048, 1920×1080, and custom ratios.

google/veo3-fast is Google’s speed-optimized AI video generation model for creating high-quality, cinematic videos with native audio from text or image prompts. It’s designed for rapid iteration, fast prototyping, and production workflows where turnaround time matters, while still preserving strong visual fidelity and prompt adherence.

This is OpenAI's general-availability realtime model, capable of responding to audio and text inputs in realtime over WebRTC, WebSocket, or SIP connections. Supports a 32,000 token context window and up to 4,096 output tokens per response (knowledge cutoff: Oct 01, 2023). Also supports image input (text/audio/image → text/audio).

Gemini Omni 1.1 Flash is Google's high-performance multimodal video generation and editing model, built on native multimodality (processing text, image, audio, and video simultaneously), conversational editing via the Interactions API, and world knowledge combining physics understanding with Gemini's broader knowledge. It generates 3-10 second video clips with audio at up to 4K (upscaled), supports text-to-video, image-to-video, first/last frame interpolation, subject and video reference conditioning, and stateful multi-turn editing and extension up to a cumulative 40 seconds. All outputs carry SynthID watermarking.

Muse Spark 1.2 is Meta's reasoning model for complex agentic tasks, accepting text, images, video, audio, and PDF documents and returning text, with a 1,048,576-token context window. Called via the Responses API, it supports cross-turn reasoning replay, tool-calling loops, built-in web search grounding, and file inputs, with configurable reasoning effort and structured JSON output. Meta reports strong results on multi-file refactors, extended debugging sessions, and whole-repository generation.

DeepSeek V3.1 is an advanced open-source hybrid AI model designed to balance powerful reasoning with high-speed efficiency. It uniquely supports two inference modes—"Think" for deep reasoning and "Non-Think" for direct, lightweight tasks—making it versatile across use cases. Built with a mixture-of-experts architecture, it scales to 685B parameters while activating only 37B per token, enabling cost-effective performance. With a 128K context window, it can handle large documents, codebases, and complex multi-step workflows. DeepSeek V3.1 is optimized for tool use, agent-based applications, and enterprise deployment through open weights and developer-friendly APIs.

Qwen3.8-27B is Alibaba's compact, deployment-friendly open-weight model from the Qwen3.8 generation, released under Apache 2.0. It is a 27.8-billion-parameter dense model using a hybrid decoder that alternates Gated DeltaNet linear attention with grouped-query full attention, with native text, image, and video understanding, a 262,144-token context window, and configurable reasoning effort toggled on or off per request. It targets coding, professional work, research, and long-horizon agentic tasks in a footprint small enough to self-host on high-end consumer hardware.

GPT‑5 is OpenAI’s flagship model, tailored for complex, multi-step reasoning and high-fidelity code generation. It combines superior world knowledge with streamlined “agentic” capabilities, enabling it to autonomously tackle tasks with minimal prompting. The model excels in logic-driven scenarios, debugging, and structured workflows—making it ideal for developers and professionals with high demands. With enhanced reasoning and smoother interaction, GPT‑5 offers users both intelligence and autonomy in one package.

GPT‑5 Mini offers a lighter and more cost-effective option, designed for everyday chat and instruction-following use cases. It retains much of GPT‑5’s reasoning power while running efficiently at lower compute costs, striking a balance between performance and affordability. Perfect for startups, pupils, or anyone mindful of resource budget, GPT‑5 Mini makes advanced capabilities more accessible. Despite its compact size, it still packs solid intelligence for routine workflows and interactions.

GPT‑5 Nano is optimized for ultra-fast performance and very low latency—ideal for real-time or embedded applications. Stripped down for speed, it excels at simple instruction-following and classification tasks, making it perfect for high-throughput API usage. With minimal overhead and rapid response times, GPT‑5 Nano is a great fit for lightweight environments or devices with limited compute capabilities. Yet, it preserves sufficient reasoning capability to remain practical across many practical scenarios.

GPT-OSS-120B is OpenAI’s most powerful open-weight model, designed for production-grade reasoning tasks with 117 billion parameters and a Mixture-of-Experts (MoE) architecture that activates 5.1 billion parameters per token. It uses a 36-layer transformer and supports long contexts (up to 128K tokens), excelling in areas like coding, math competitions, and health applications. The model offers strong tool-use capabilities (e.g., browsing, function calling) and performs on par with o4-mini in core reasoning benchmarks.

GPT-OSS-20B is a lightweight, efficient open-weight model optimized for low-latency, on-device, and edge deployments. With 21 billion parameters and a similar MoE structure as its larger sibling, it supports 128K token contexts while running comfortably on 16GB consumer hardware. It matches or beats o3-mini on standard tasks and excels in coding, math, and health-related reasoning.

Claude Opus 4.1, a direct replacement for Opus 4, excels in real-world coding and agentic tasks with enhanced performance and precision. It autonomously handles complex, end-to-end development tasks, adapting to user styles while maintaining high-quality output, particularly in frontend code generation and intricate logic. With superior long-horizon task management and problem-solving, it serves as an ideal virtual collaborator for sustained reasoning and multi-step actions. Opus 4.1 also boosts AI agent capabilities, enabling accurate execution of complex tasks and shining in agentic search, research, content creation, and context management for comprehensive insights and high-quality summarization.

Lucid Origin is Leonardo AI's versatile text-to-image model optimized for exceptional prompt adherence, vibrant Full HD outputs, and diverse visual styles from photorealistic to stylized illustration.

Qwen3.8-Max is Alibaba's flagship sparse mixture-of-experts model, with 2.4 trillion total parameters and approximately 95 billion active per forward pass, accepting text, image, and video input and returning text with a 1M-token context window. Hybrid thinking mode is enabled by default with three reasoning-effort tiers, and pricing is flat across the full context window with no long-prompt surcharge. Target workloads include software engineering, long-horizon agentic work, office productivity, and visual tasks such as turning screenshots or design files into working pages.

Qwen-Image is an image generation foundation model in the Qwen series that achieves significant advances in complex text rendering and precise image editing. It supports parallel CFG and LoRA merging (up to 3 LoRAs), configurable inference steps, and an optional turbo mode for faster generation using optimized settings (10 steps, CFG=1.2).

Z-Image is the foundation model of the Z-Image family, engineered for good quality, robust generative diversity, broad stylistic coverage, and precise prompt adherence. It is a 6B-parameter text-to-image model built for efficient, high-fidelity image generation with strong prompt following capabilities across a wide range of visual styles.

Z-Image Turbo is a super fast text-to-image model of 6B parameters developed by Tongyi-MAI. It compresses inference to as few as 1-8 diffusion steps (default 8), delivering high-volume, low-cost image generation at $0.005 per megapixel while maintaining strong prompt adherence. Optimized for rapid prototyping, content variation testing, and high-throughput asset generation, with optional prompt expansion to enrich brief prompts automatically.

HappyHorse 1.1 is Alibaba's closed-weight cinematic video generation model, producing clips of 3-15 seconds with synchronized audio and multilingual lip-synced dialogue generated in a single pass. It generates video from a text prompt, a single starting image, or a set of reference images, and supports video editing of existing clips, at output resolutions of 480p, 720p, and 1080p. It improves on HappyHorse 1.0 with stronger prompt adherence, smoother motion, and more consistent characters across frames, and previously topped the Artificial Analysis video arena anonymously before Alibaba was revealed as its maker.

Muse Spark 1.1 is Meta's first paid agentic reasoning model, accepting text, image, video, and PDF input and returning text, with a 1,048,576-token context window and 131,072 max output tokens. It thinks before it answers with reasoning tokens billed as output, actively manages and compacts its own context on long-running workflows, and exposes native tool calling, structured output, parallel tool calling, and built-in web search grounding with inline citations via the Responses API.

Lucid Realism is a specialized text-to-image model from Leonardo AI optimized for hyperrealistic, cinematic photorealism with exceptional detail, natural lighting, and minimal generation errors.

Qwen3-Embedding-8B is the largest model in the Qwen3 Embedding series, purpose-built for text embedding and retrieval tasks. It ranks No.1 on the MTEB multilingual leaderboard, supports 100+ languages, and produces embeddings up to 4096 dimensions with support for user-defined output dimensions (Matryoshka Representation Learning) from 32 to 4096. The model is instruction-aware, allowing custom task instructions to be prepended to inputs for improved retrieval performance, and supports a 32K token context length.

FLUX.1 Kontext [pro] is an advanced image generation and editing model from Black Forest Labs that supports both text prompts and reference images for in-context modifications.

google/veo3 is Google DeepMind’s advanced AI video generation model that creates high-quality, cinematic videos with synchronized audio from text prompts or image references. It’s designed for realistic motion, strong prompt adherence, and audiovisual consistency, making it useful for short-form storytelling, concept visualization, and professional content creation.

Sarvam: Bulbul V2 is a text-to-speech (TTS) model from Sarvam AI, delivering real-time, natural-sounding speech in 11 Indian languages at a fraction of global costs. It builds on prior versions with key enhancements including improved audio quality, 30+ speaker voices (expanded from six distinct voices in v1 for diverse use cases like professional or conversational tones), and adjustable speech speed from 0.5x to 2.0x for customized delivery.

Gemini 2.5 Pro is Google's premiere thinking model built for enterprise-grade applications requiring sophisticated reasoning. It delivers top performance on critical benchmarks, featuring built-in reasoning capabilities that enable it to plan and analyze before responding to complex queries. The model maintains high accuracy across STEM domains with native multimodal processing and an extensive context window. This preview release allows early access to Google's most advanced model, which excels in code development, scientific problem-solving, and multi-step logical tasks while maintaining human-aligned responses. The preview designation indicates ongoing optimization before general availability.

A Mixture-of-Experts model that activates only 3.3B of its 30.5B parameters per forward pass. Offers dual thinking modes: a detailed step-by-step reasoning mode for complex problems and a faster non-thinking mode for simpler queries. Despite its small active parameter count, it outperforms QwQ-32B that has 10 times more activated parameters.

A versatile dense model offering powerful reasoning abilities paired with efficient dialogue processing. Developed by Alibaba, it supports extensive 32K context windows (extendable to 131K), features both thinking and non-thinking modes, and excels in coding, mathematics, and multilingual support across 119 languages and dialects.

Qwen3-32B is a dense 32.8B parameter causal language model from the Qwen3 series, optimized for both complex reasoning and efficient dialogue. It supports seamless switching between a "thinking" mode for tasks like math, coding, and logical inference, and a "non-thinking" mode for faster, general-purpose conversation. The model demonstrates strong performance in instruction-following, agent tool use, creative writing, and multilingual tasks across 100+ languages and dialects. It natively handles 32K token contexts and can extend to 131K tokens using YaRN-based scaling.

Gemini 2.5 Flash is Google's hybrid reasoning model offering an optimal balance between performance, cost, and latency. It features full thinking capabilities that can be toggled on/off or fine-tuned using a thinking budget parameter (0-24576 tokens). This model excels at complex tasks requiring multi-step reasoning while maintaining the efficiency needed for high-volume and real-time applications. It outperforms comparable models in reasoning benchmarks while maintaining lower cost and latency profiles, making it ideal for enterprise deployments requiring both deep analysis and scalability.

OpenAI o4-mini-high is the enhanced reasoning variant of o4-mini with reasoning_effort permanently set to "high" for maximum problem-solving capabilities. Released in April 2025, this configuration prioritizes thorough analysis and step-by-step reasoning over processing speed, making it ideal for complex STEM problems, mathematical proofs, and detailed code generation tasks. While maintaining the same model architecture as standard o4-mini, this variant allocates additional computational resources to reasoning processes, resulting in higher accuracy on complex problems with a corresponding increase in token usage and processing time. The model demonstrates exceptional performance on sophisticated benchmarks including AIME and achieves impressive results on visual reasoning tasks. This specialized configuration is particularly valuable for applications where solution quality and reasoning thoroughness outweigh speed considerations, such as educational tools, research assistance, and complex problem-solving systems requiring maximum accuracy.

OpenAI o3 is an advanced reasoning model, setting new standards in coding, math, science, and visual understanding. It outperforms on benchmarks like Codeforces, SWE-bench (without custom scaffolding), and MMMU. Ideal for complex, nuanced problems, o3 excels at visual tasks and reduces major errors by 20% compared to o1 in real-world evaluations. It’s especially strong in programming, business, and creative ideation, with early users praising its ability to generate and critically assess novel ideas in biology, math, and engineering.

OpenAI o4-mini combines exceptional efficiency with sophisticated reasoning capabilities in a compact model from the o-series. Released in April 2025, this streamlined solution maintains powerful multimodal and agentic capabilities while significantly reducing computational requirements. The model achieves remarkable benchmark performance, including near-perfect scores on AIME with Python (99.5%) and competitive results on SWE-bench, outperforming its predecessor o3-mini and approaching the larger o3 model in several domains. Despite its optimized size, o4-mini excels in STEM tasks, visual problem-solving (MathVista, MMMU), and code editing applications. Its refined reinforcement learning architecture enables sophisticated capabilities like tool chaining, structured output generation, and multi-step task completion with minimal latency—often resolving complex problems in under a minute, making it ideal for high-throughput scenarios prioritizing both speed and quality.

GPT-4.1 is OpenAI's latest flagship model showcasing significant advances in coding, instruction following, and long-context understanding. It processes up to 1 million tokens (approximately 750,000 words) in a single context window, with knowledge updated to June 2024. Performance benchmarks demonstrate substantial improvements over GPT-4o, including 54.6% completion on SWE-bench Verified coding tasks (21.4% increase) and 38.3% on MultiChallenge instruction following (10.5% increase). The model excels at real-world software engineering with reduced extraneous edits (from 9% to 2%), superior code exploration capabilities, and enhanced long-document comprehension. GPT-4.1 provides improved agentic reliability while maintaining a lower price point than its predecessors, making it particularly valuable for development environments, document analysis, and enterprise knowledge systems.

GPT-4.1 Mini delivers GPT-4o-level performance with significantly improved efficiency metrics across latency and cost. This mid-sized model maintains the full 1 million token context window of its larger counterpart while achieving impressive benchmark scores: 45.1% on hard instruction evaluations, 35.8% on MultiChallenge, and 84.1% on IFEval. Despite its reduced parameter count, GPT-4.1 Mini demonstrates robust coding capabilities (31.6% on Aider's polyglot diff benchmark) and strong vision understanding. The model reduces latency by nearly half and costs 83% less than GPT-4o, making it ideal for interactive applications with tight performance constraints, high-throughput services, and cost-sensitive enterprise deployments requiring responsive AI capabilities without compromising on quality.

GPT-4.1 Nano is OpenAI's most efficient model optimized for maximum speed and minimum cost in the GPT-4.1 family. Released in April 2025, this compact model maintains the full 1 million token context window while delivering impressive benchmark performance: 80.1% on MMLU, 50.3% on GPQA, and 9.8% on Aider polyglot coding—outperforming even GPT-4o mini. At just $0.10 per million input tokens and $0.40 per million output tokens, it represents OpenAI's most affordable option. The model excels in latency-critical applications including classification, autocompletion, information extraction, and high-throughput document processing where efficiency and cost considerations are paramount.

Grok 3 Mini Beta is xAI's lightweight reasoning model designed specifically for quantitative problem-solving with transparent thinking processes. Unlike conventional AI systems that generate immediate answers, this model implements explicit reasoning steps before responding, making it particularly effective for mathematics, puzzles, and logical reasoning tasks. The model defaults to low reasoning effort but can be configured with increased reasoning capacity (reasoning: { effort: high }) for more complex problems. With accessible thinking traces, users can examine the model's step-by-step problem-solving approach, providing valuable insights into its reasoning process. This beta release offers early access to xAI's reasoning capabilities through dual endpoint options: a standard endpoint optimized for quality and a fast endpoint prioritizing throughput for latency-sensitive applications.

Grok 3 Beta is xAI's flagship model featuring specialized domain expertise in finance, healthcare, law, and scientific fields. This enterprise-focused solution delivers exceptional performance in data extraction, code generation, and text summarization tasks. Released in 2025, the model particularly excels in structured tasks and achieves superior results on challenging benchmarks like GPQA, LCB, and MMLU-Pro, consistently outperforming the smaller Grok 3 Mini even when the latter operates at high thinking levels. Grok 3's sophisticated knowledge architecture makes it particularly valuable for professional and technical applications requiring deep contextual understanding in regulated industries. The model offers dual deployment options: a standard endpoint optimized for quality and a fast endpoint prioritizing throughput, allowing users to select the appropriate performance profile based on their specific application requirements.

google/veo2 is Google DeepMind’s AI video generation model for creating high-quality, realistic videos from text prompts or image references. It’s built for strong motion quality, cinematic detail, and accurate prompt following, making it useful for storyboarding, concept visualization, and short-form video generation.

Deepgram Nova-3 is a fast, accurate speech-to-text model served through Telnyx's OpenAI-compatible transcription endpoint. It supports English variants (en, en-US, en-GB, en-AU, en-NZ, en-IN) and only accepts mp3 or wav audio files. Deepgram-specific options such as smart formatting, punctuation, diarization, utterance segmentation, and numeral conversion are configurable via model_config, and word-level timestamps with speaker labels are returned when enabled.

Sonar Reasoning Pro is a high-performance reasoning model designed to tackle complex topics that require advanced multi-step thinking and deep content understanding. With a context length of 128k, it combines enhanced Chain-of-Thought (CoT) reasoning with a powerful information retrieval architecture to deliver well-structured, thoughtful responses. It provides twice as many citations on average compared to the standard Sonar Reasoning model, ensuring greater transparency and grounding. Ideal for tasks like competitive product analysis, exploring intricate scientific subjects, or creating detailed travel plans, Sonar Reasoning Pro excels in scenarios that demand both reasoning and reliable, well-sourced information.

Sonar Pro is an advanced non-reasoning search model designed for handling complex, multi-step queries that demand deeper content understanding and high citation accuracy. With a large context length of 200k, it excels at retrieving and synthesizing information from extensive text sources. Sonar Pro delivers in-depth answers supported by twice as many citations as the standard Sonar model, thanks to its enhanced information retrieval architecture. Optimized for multi-step tasks, it’s particularly useful for applications such as academic literature reviews, competitor and industry research, and generating curated content like restaurant catalogs with reviews. Its ability to surface comprehensive, well-cited responses makes it a powerful tool for users needing detailed and reliable information.

Sonar Deep Research is a powerful deep research and reasoning model designed for conducting thorough investigations and generating expert-level insights. With a context length of 128k, it excels at synthesizing information from hundreds of sources to produce highly detailed, structured reports. Ideal for exhaustive research tasks, Sonar Deep Research offers in-depth subject analysis and is capable of crafting content with precision and authority. Common use cases include writing white papers for industry thought leadership, developing comprehensive go-to-market (GTM) strategies, and creating advanced educational material for universities or training programs. It's the ideal tool for users who need depth, clarity, and rigor in their research output.

State-of-the-art Gemini embedding model (3072-dim) optimised for retrieval and semantic similarity.

Omni Moderation (Latest) is a state-of-the-art moderation model designed for real-time classification of unsafe, offensive, or policy-violating content. It supports text moderation across multiple languages and content types, including user messages, document uploads, code, and social media content. The model is optimized for low false positives and fast ruling in both API and streaming workflows.

Seedance 1.0 Pro is the full version of the Seedance model (ByteDance) with professional-level capabilities. Like the Lite version, it creates 1080p multi-shot videos from text or images, but with enhanced storytelling features. Pro supports wide dynamic motion ranges for smooth large-scale movements and native multi-shot consistency (stable subjects and style across cuts). It also interprets diverse visual styles accurately. In benchmarks, Seedance Pro shows very high prompt adherence, motion quality, and aesthetic scores. It's intended for creating polished, multi-scene videos in applications like advertising or short films.

Vidu 2.0 is an improved model (early 2024) that focuses on image-to-video generation. It adds support for more sophisticated animations and produces higher-quality output than Vidu 1.x. The model is better at interpreting detailed visual cues, allowing more intricate camera moves and character actions. It typically outputs at HD resolution. Compared to its predecessor, Vidu 2.0 yields smoother, more polished videos with richer detail. It's commonly used for bringing concept art or still images to life in an animated clip.

Vidu Q1 (2025) is the newest model from Vidu AI (ByteDance) for text/image to video. It produces 1080p clips with outstanding realism; reviewers call its outputs "arguably the most lifelike" seen to date. The model features very sharp visuals, smooth transitions, and detailed scene composition. It excels at cinematic lighting and keeps characters consistent, even when following complex emotional or action-packed prompts. Built on an intuitive interface, Vidu Q1 balances ease of use with high-end results. In short, it specializes in generating high-fidelity videos with strong narrative coherence and rich visual effects.

Pika 2.2 (early 2025) is a versatile video model supporting text-to-video and image-to-video. It generates dynamic 10-second clips at 1080p with improved realism and prompt responsiveness. New features include keyframe transitions (Pikaframes) that smoothly interpolate between start/end images, and various AI-powered effects (Pikaffects, Pikadditions, etc.) for creative control. Pika 2.2 delivers cinematic motion and camera effects, though very complex prompts can still produce occasional artifacts. Overall, it represents a major upgrade with higher resolution and fluidity over Pika 2.1, aimed at content creators and social media applications.

Kling 2.0 (2024) was Kling's image/video model aimed at cinematic output. It introduced higher resolution outputs and smoother camera movements than Kling 1.x. The model generated realistic motion and stable scenes, with a focus on high-quality visual polish. Compared to the later 2.1, Kling 2.0 was a solid general-purpose model but had fewer advanced features (e.g. no multi-frame generation). It supported both image and text prompts for 720p/1080p animations. In effect, Kling 2.0 provided film-like AI videos that set the stage for even more advanced models.

Kling 1.6 (Dec 2024) was a major upgrade over Kling 1.5, reportedly boosting performance by ~195%. It features much better prompt understanding and physics, resulting in very natural movements and facial expressions in the generated videos. The image-to-video quality is significantly enhanced: videos are more vibrant and consistent, with refined lighting and shadows. Kling 1.6 offers dual modes (Standard and Professional) to balance speed vs. quality. Notably, these improvements were delivered with no extra cost to users, making it a highly attractive option for free AI video generation.

Runway Gen-3 was Runway's early video model (circa 2023) for professional animation tasks. It can animate still images or generate text-to-video clips up to 10 seconds long. Gen-3 outputs were typically 720p (with upscaling options) and aimed for cinematic style. It emphasized creative control, allowing prompts to specify actions, camera moves, and visual styles. Users praised its consistency in characters and environments, though it could struggle with very complex scenes. In practice, Gen-3 served as a reliable workhorse for creative industries, producing high-quality, animated content and setting the stage for its successors.

Runway Gen-4 Turbo is optimized for speed and cost. By default it outputs 720p video but supports 4K upscaling for higher quality. Turbo preserves much of Gen-4's motion realism—handling dynamic lighting and depth convincingly—while sacrificing some fine details in very complex scenes. This makes it ideal for rapid prototyping and scenarios where quick turnaround is needed. Overall, Gen-4 Turbo delivers visually impressive 720p clips with fluid action and is faster (and cheaper) than the original Gen-4, at the cost of slight quality tradeoffs.

Kling 2.1 is Kuaishou's advanced image-to-video model (2025). It transforms reference images (and optional text) into high-definition cinematic videos using a 3D spatiotemporal attention architecture. The model simulates real-world physics and delivers very natural, intricate motion and lifelike characters. It can produce multiple video variants from the same prompt and uses AI-assisted prompting to help users craft detailed instructions. In comparative tests, Kling 2.1 rivals top closed-source competitors. Overall, it marks a significant step up in realism, coherence, and prompt-following over Kling 2.0.

Kling 2.1 Master is the premium version of Kling's 2.1 model, designed for professional use. It delivers top-tier image-to-video performance with exceptionally fluid motion and cinematic visuals. The model provides smoother playback and more dramatic action scenes compared to lower tiers. It offers very precise prompt adherence, allowing complex instructions to produce coherent, high-quality video. The high output quality comes with increased compute cost, but users gain a very polished, filmic look. In summary, Kling 2.1 Master produces advanced, dynamic animations with excellent detail and strict alignment to input prompts.

Pollo 1.6 is Pollo AI's flagship video model, optimized for efficiency. Described as "better, cheaper, and faster," targeting high-quality cinematic outputs, it supports both image-to-video and text-to-video inputs. In practice, 1.6 delivers more realistic and creative animations than previous Pollo versions. It emphasizes speed and cost-effectiveness, making it accessible for a wide range of users. Overall, Pollo 1.6 aims to balance visual fidelity with fast generation, serving as a versatile general-purpose model.

ACE-Step is an open-source model built to address the shortcomings of previous music generation systems by combining diffusion techniques, Sana’s Deep Compression AutoEncoder (DCAE), and a lightweight linear transformer. The ace-step/prompt-to-audio functionality is an interface within the ACE-Step music generation model that enables users to create audio directly from text prompts. This architecture allows it to produce high-quality, musically coherent audio efficiently and with a high degree of control. Users can input various prompt types -- including short tags, descriptive phrases, or full lyrics -- and the model interprets these inputs to generate corresponding music. With features like lyric editing and detailed control over song structure and style, the prompt-to-audio tool offers a powerful, flexible way to create music using natural language.

GPT-image-1 is OpenAI's latest multimodal image generation model, released in April 2025, and represents a significant advancement over previous models like DALL-E. It can generate high-resolution images—up to 4096×4096 pixels—from natural language prompts, with improved fidelity and consistency in handling complex scenes and detailed instructions. Key features include robust text-to-image generation, reliable text rendering within images, and advanced editing capabilities such as image-to-image transformation and inpainting, allowing users to upload and modify existing images using text prompts.

Llama-4-Scout-17B-16E-Instruct is a multimodal, instruction-tuned language model developed by Meta, designed to handle long-context tasks efficiently. It features 17 billion active parameters with a total of 109 billion parameters across 16 experts, utilizing a mixture-of-experts (MoE) architecture. The model supports a context window of up to 10 million tokens, making it suitable for applications requiring extensive context, such as multi-document summarization and large codebase exploration.

Llama 3.1 8B Instruct is Meta's compact, efficiency-focused language model from the Llama 3.1 family. At just 8 billion parameters, this instruction-tuned variant delivers impressive performance with minimal computational requirements, making it ideal for resource-constrained environments and latency-sensitive applications. Despite its smaller size, the model maintains competitive performance in human evaluations against much larger models, particularly excelling in straightforward dialogue, content generation, and common instruction-following tasks. The model features improved inference speed and reduced deployment costs compared to larger variants while maintaining the core capabilities that make Llama models effective. It's particularly well-suited for mobile applications, edge devices, and scaled deployments where efficiency and performance balance are critical considerations.

OpenAI o3-mini (January 2025 version) is a specialized reasoning model optimized for STEM domains while maintaining cost efficiency. This model features adjustable reasoning capabilities and demonstrates exceptional performance in science, mathematics, and coding tasks, with expert evaluators preferring its responses 56% of the time over previous versions and noting a 39% reduction in major errors on complex questions. Even at medium reasoning settings, o3-mini matches the larger o1 model's performance on challenging evaluations like AIME and GPQA while maintaining significantly lower latency and cost profiles. The model supports developer-focused features including function calling, structured outputs, and streaming capabilities, though it lacks vision processing functionality. This text-only design focuses computational resources on reasoning quality, making it particularly valuable for technical applications requiring accurate problem-solving without multimodal requirements.

OpenAI o3-mini-high is the enhanced reasoning variant of o3-mini with reasoning_effort permanently set = high for maximum problem-solving capabilities. Released in January 2025, this specialized configuration prioritizes thorough analysis and step-by-step reasoning over processing speed, making it ideal for complex STEM problems requiring in-depth analysis. While maintaining the same core architecture as standard o3-mini, this variant allocates additional computational resources to reasoning processes, resulting in higher accuracy on challenging problems in science, mathematics, and coding domains. The model maintains all developer-focused features of the standard version, including function calling, structured outputs, and streaming capabilities, while focusing its additional reasoning capacity on problem complexity rather than multimodal processing. This configuration is especially valuable for applications where solution quality and reasoning thoroughness outweigh processing speed considerations, such as educational tools, research assistants, and technical problem-solving systems requiring maximum accuracy in areas demanding rigorous analytical thinking.

A 24B parameter model optimized for speed and efficiency while rivaling models 3x its size like Llama 3.3 70B. Achieves 81% accuracy on MMLU benchmarks with impressive 150 tokens/second throughput on consumer hardware. Released under Apache 2.0 license, it serves as an excellent base for developing advanced reasoning capabilities.

Sonar is a lightweight, cost-effective non-reasoning search model built for delivering fast, grounded answers through real-time web search. With a context length of 128k, it is optimized for quick lookups and straightforward Q&A tasks, making it ideal for users who need immediate, reliable information without extensive processing. Sonar provides answers backed by citations, ensuring transparency while maintaining speed and low cost. Typical use cases include summarizing books, TV shows, or movies, looking up definitions or quick facts, and browsing up-to-date content across categories like news, sports, health, and finance. Its efficiency and accuracy make it well-suited for everyday information needs.