FastRoutervs.Portkey
Pre-acquisition, Portkey was the closest competitor to FastRouter on feature depth. Palo Alto Networks announced its intent to acquire Portkey on April 30, 2026; the product is being integrated into Prisma AIRS. Existing customers continue to be supported. The buying decision for the next 12–24 months changed.
By Ritesh Prasad~14 min readShort version
The quick decision
| If you are... | Use... | Why? |
|---|---|---|
| Currently on Portkey, evaluating options | FastRouter | Independent roadmap; comparable feature surface; deeper eval layer. |
| Need the broadest guardrail integration catalog | Portkey | 60+ built-ins, plus Patronus, Aporia, Bedrock and Azure native guardrails. |
| Already a Palo Alto Networks / Prisma AIRS customer | Portkey | Single vendor, single procurement, single security review. |
| Choosing a gateway for the next 12–24 months | FastRouter | Avoid the PAN integration window. Built-in evals, GEPA, MCP credential vaulting. |
At a glance: Key metrics
| Metric | FastRouter | Portkey |
|---|---|---|
| Roadmap owner | Independent, developer-first | Palo Alto Networks → Prisma AIRS |
| Model catalog | Major frontier + open providers | 1,600+ models |
| Routing strategies | 7 (incl. AI Auto Router) | Fallback + conditional + load balance |
| Built-in evals | Smart + Auto + GEPA + Video | Batch evals + feedback API |
Feature matrix
| Capability | FastRouter | Portkey |
|---|---|---|
| Roadmap ownership | Independent | Palo Alto Networks (Prisma AIRS) |
| Model catalog | Major frontier + open | 1,600+ models |
| Category-based routing | First-class strategy | Via conditional routing on metadata |
| Eval-driven Auto Router | Picks per request from your eval scores | Not supported |
| GEPA prompt optimization | Proprietary | Not supported |
| Guardrail marketplace | Built-in PII redaction, smaller catalog | 60+ built-ins, Patronus, Aporia, Bedrock/Azure |
| Open-source self-hosted gateway | Managed only | MIT, ~122KB, self-hostable |
| Semantic caching | Across plans | Enterprise tier only |
Acquisition context
Palo Alto Networks announced its intent to acquire Portkey on April 30, 2026. The deal closes in PAN's Q4 fiscal 2026, with Portkey integrating into Prisma AIRS (PAN's enterprise AI security platform). Existing customers continue to be supported.
The forward-looking risk: roadmap priorities tilt toward the security-buyer (CISO orgs, compliance teams) over the developer-first audience Portkey grew up serving. This is a common post-acquisition pattern in this category, not a guarantee.
Deep dive: Routing & evals
1. Routing strategy
Portkey takes a configuration-as-code approach (JSON Configs) with fallback chains, conditional routing on request metadata, and weighted load balancing. FastRouter ships the same primitives, then adds category routing as a first-class strategy (not emulated through conditional rules), highest-throughput as an explicit sort axis, and an AI Auto Model Router driven by your production eval scores.
2. The eval workflow gap
Portkey's eval primitives — a Feedback API plus batch evaluations for reliability, grounding, and bias testing — fit offline workflows. FastRouter runs Smart and Automatic Evaluations on live traffic, ships GEPA for prompt optimization, supports video evals, and feeds the scores back into the routing layer so they influence the next request.
Final decision tree
Currently on Portkey and it's working? → Stay for now. Re-evaluate in 6–12 months as PAN roadmap signals clarify. Run a parallel FastRouter audit for comparison.
Choosing a gateway today for the next 12–24 months? → FastRouter. Independent roadmap, comparable features, deeper eval layer.
Specific 3rd-party guardrail requirements (Patronus, Aporia)? → Portkey. The marketplace breadth is genuinely deeper.
Already a Palo Alto Networks customer? → Talk to PAN about the Prisma AIRS bundle.
Side-by-side
The full feature breakdown
✓ supported, ✗ not supported, ◑ partial.
| Capability | FastRouter Independent | Portkey PAN / Prisma AIRS |
|---|---|---|
| Roadmap ownership | Independent, developer-first | Palo Alto Networks → Prisma AIRS |
| Model catalog | Major frontier + open providers | 1,600+ models — broadest in the managed category |
| Routing strategies | 7: category, priority, lowest latency, lowest price, highest throughput, weighted, AI Auto | Fallback (with custom status codes), conditional routing on metadata, weighted load balancing |
| AI Auto Model Router | Picks per request from cost, latency, and your eval scores | Not supported |
| Category-based routing | First-class strategy | ◑Achievable via conditional routing on metadata |
| Smart / Automatic Evaluations | Live production traffic | ◑Batch evals + feedback API; not continuous |
| GEPA prompt optimization | Proprietary | Not supported |
| Video evaluations | Compare models on video inputs | Not supported |
| MCP gateway | + credential vaulting | Works with Claude Desktop, Cursor, VS Code, any MCP client |
| Prompt management (library, versions) | Supported | Mature prompt library with versioning |
| Semantic caching | Across plans | ◑Enterprise tier only |
| Guardrail marketplace | ◑Built-in PII redaction, smaller third-party catalog | 60+ built-ins, Patronus, Aporia, Bedrock/Azure guardrails |
| Workspaces / virtual keys / budgets | Workspace-level kill-switches | Mature; per-developer/team caps |
| SSO (Okta, Azure AD) | Supported | Supported |
| SOC 2 Type 2 / ISO 27001 / HIPAA | Supported | Supported (Enterprise) |
| Open-source self-hosted gateway | Not supported | MIT, ~122KB, self-hostable |
| 7-day passive audit | Zero code changes | Not supported |
1,600+ model catalog, broader guardrail marketplace (Patronus, Aporia, Bedrock/Azure native), and an MIT-licensed self-hostable Gateway. If any of those are decisive for your stack, Portkey is still a credible answer despite the acquisition.
Acquisition context
What the PAN deal changes
Palo Alto Networks framed the acquisition as bringing Portkey's "unified control plane for AI agents" into Prisma AIRS, PAN's enterprise AI security platform. The press release emphasizes agent governance, AI threat protection, and unifying AI security under PAN's enterprise sales motion.
That framing tells you what's about to get prioritized. Enterprise security buyers (CISOs, compliance teams) are a different audience than the AI engineering teams Portkey grew up serving. Practically:
Pricing tends to migrate toward enterprise contract motion. Self-serve and small-team plans often quietly atrophy.
Roadmap tilts toward security-buyer priorities: agent risk classification, prompt-injection defenses, audit and reporting depth.
Iteration speed usually slows as engineering teams integrate with the parent's systems (auth, billing, observability, security review).
Open-source attention on the Portkey Gateway repo may diverge from cloud SaaS priorities.
None of this is a critique of PAN — Prisma AIRS is a real product. It's a fit shift for teams whose primary need is a developer-first, fast-iterating LLM gateway.
Nothing breaks because of the announcement. Existing customers continue to be supported and the deal closes in PAN's Q4 fiscal 2026. The realistic move: stay on Portkey for current workloads, evaluate alternatives in parallel for the next 12–24 months, and re-decide when PAN's roadmap signals get concrete.
Routing & gateway
Configuration-as-code vs strategies
Portkey's model: Configs
Configuration-as-code. You define a Config — a JSON document declaring fallback chains, conditional routing rules (route on request metadata), and weighted load balancing — save it, reference its Config ID in requests. Conditional routing is powerful: route Claude for "support_tickets" and GPT for "code_generation" by inspecting metadata at the edge. Fallbacks compose, so you can build sophisticated multi-target failover trees.
FastRouter's seven strategies
Same primitives plus three that Portkey doesn't ship: category-based routing as a first-class strategy (rather than emulated via conditional rules), highest-throughput as an explicit optimization axis, and the AI Auto Model Router that picks per request from cost, latency, and quality signals fed by the eval layer.
On gateway primitives — caching, retries, timeouts, multimodal, MCP — both are mature. Portkey's semantic caching is gated to Enterprise; FastRouter ships it across plans. Portkey's MCP support is broad and well-documented; FastRouter adds credential vaulting on top.
Portkey routes between providers using configs you write. FastRouter routes between providers and between models — using signals it gathers automatically.
Evaluations & optimization
Batch evals vs continuous evals
Portkey's eval story
Two primitives: a Feedback API for attaching scores (-10 to +10, optional weight) to traces by trace ID, and batch evaluations for offline reliability, grounding, and bias testing. Data feeds an evals dashboard alongside Analytics. Clean implementation that fits offline eval workflows.
FastRouter's continuous story
Smart Evaluations score live calls automatically. Automatic Evaluations sample real traffic to benchmark competing models in the background. GEPA evolves prompt and model combinations toward Pareto-optimal cost/quality. Video evaluations extend this to a content type no other gateway in the category currently supports.
The practical difference: Portkey expects you to bring evaluation logic to it (datasets, scoring functions, feedback loops). FastRouter generates the signals continuously and feeds them back into routing.
Portkey ships evals as a tool you wield. FastRouter ships evals as a substrate that improves the gateway's behavior on its own.
Governance & security
Compliance parity, different threat-model emphasis
On compliance certifications — SOC 2 Type 2, GDPR, HIPAA, ISO 27001 — Portkey and FastRouter are roughly at parity. Both ship workspace governance, virtual keys with budgets and rate limits, RBAC, SSO via Okta and Azure AD, and audit logs.
Portkey's guardrail catalog is broader
60+ built-ins plus Patronus AI (hallucination detection, RAG-specific), Aporia (data leak prevention, policies), and integrations with Bedrock and Azure native guardrails. If your security team has specific 3rd-party guardrail requirements, this is a real Portkey advantage.
FastRouter's MCP credential vault
Designed for agent workloads where the gateway is the only place that ever holds raw provider keys. Agents and tool callers never see them — they call FastRouter, FastRouter injects the credential server-side. Closes the exfiltration vector where a compromised MCP tool can leak provider credentials.
Same compliance shelf. Portkey leads on third-party guardrails; FastRouter leads on agent-key handling.
Roadmap risk
How acquisitions typically play out
Common pattern when a fast-moving developer-tool company gets acquired by a larger enterprise platform vendor:
Apr 30, 2026 — PAN announces intent to acquire Portkey. Press release positions Portkey as the agent control plane inside Prisma AIRS. Existing customers continue to be supported.
Now → close (Q4 PAN fiscal 2026) — Operations continue. Portkey's product surface unchanged for current customers. Roadmap announcements typically pause during the integration planning window.
~6 months post-close — Integration with parent's systems begins (auth, billing, support, security review). Iteration on standalone product features typically slows.
~12 months post-close — Roadmap re-orients toward parent's buyer. Features tilt toward security/compliance reporting and enterprise governance. Self-serve and small-team plans often atrophy.
~18–24 months post-close — Standalone product often becomes a feature inside the parent's bundle. Sometimes the brand survives; the standalone roadmap usually doesn't.
This is a base-rate observation, not a prediction. Some acquired products thrive and ship faster post-acquisition. Worth weighing consciously.
Honest take
When each one wins
When Portkey is the better pick
You need the broadest guardrail integration catalog
- Patronus, Aporia, plus 60+ built-ins
- Bedrock and Azure native guardrails wired in
- Compliance teams have specific 3rd-party requirements
You want the MIT-licensed gateway as a fallback
- Open-source Gateway (~122KB) you can self-host if needed
- Optionality if cloud product changes post-acquisition
- Code visibility for security review
You're already a Palo Alto Networks customer
- Procurement, billing, and security review already in place
- Prisma AIRS bundle may make commercial sense
- Single vendor relationship for AI security stack
You need raw model breadth (1,600+)
- Largest managed-gateway catalog
- Niche/long-tail providers covered
- Easier to land on a single API for everything
When FastRouter is the better pick
You're choosing for the next 12–24 months
- Independent, developer-first roadmap
- No integration window to wait through
- No bet on what Prisma AIRS becomes
You want continuous evals, not batch evals
- Smart + Automatic Evaluations on live traffic
- GEPA prompt optimization runs continuously
- Eval signals feed AI Auto routing decisions
You're running serious agentic workloads
- MCP credential vaulting — agents never see raw keys
- AI Auto Model Router selects per request
- Workspace-level budget kill-switches
You want video evaluations
- Compare model quality on video inputs
- No other gateway in the category supports this
- Useful for multimodal product workflows
How to choose
The decision tree
You're currently on Portkey and the product is working
Stay for now. Re-evaluate in 6–12 months once PAN's roadmap signals get concrete. Run a parallel FastRouter audit so you have a comparison when it's time.
You're choosing a gateway today for the next 12–24 months
Default to FastRouter. Independent roadmap, comparable feature parity, deeper eval and optimization layer.
Your security org has specific 3rd-party guardrail requirements (Patronus, Aporia)
Use Portkey. The marketplace breadth is genuinely deepest in the category.
You're already a Palo Alto Networks customer and want a single vendor
Talk to PAN about the Prisma AIRS bundle. Single procurement, single security review.
You need an MIT-licensed self-hostable gateway as a fallback
Pin Portkey Gateway (the OSS repo) at a known-good version. Use it as a contingency alongside a managed product.
Things people ask before they switch
No. PAN has stated existing customers continue to be supported and the deal closes in PAN's Q4 fiscal 2026. Day-to-day operations are unchanged. The risk is forward-looking — what gets built (or doesn't) over the next 12–24 months.
Both are OpenAI-compatible. The bulk of migration is endpoint and key changes. Portkey Configs (fallback, conditional routing) port to FastRouter strategies — usually a 1:1 mapping with category-based routing replacing some conditional rules. Virtual keys map to FastRouter API keys. We help with cutover for production workloads.
Yes — for what it does. The OSS Gateway (~122KB, MIT) is fast and solid. It's the gateway primitive only — observability, governance, prompt management, and evals are cloud-only. If your need is "give me a fast multi-provider proxy I can self-host," the OSS Gateway alone may be enough.
It's real and broader than FastRouter's. If you specifically need a long-tail model that only one provider hosts, Portkey is more likely to have it on day one. For the top 30 models that drive ~95% of production traffic in 2026, both cover them.
Genuinely broader than FastRouter's today. 60+ built-ins, plus Patronus AI (hallucination detection, RAG-specific), Aporia (data leak prevention, policies), and integrations with Bedrock and Azure native guardrails. If your security team has specific 3rd-party guardrail requirements, this is a real Portkey advantage.
Yes — it's on the roadmap. We ship built-in PII redaction and prompt-injection guardrails today, and the third-party integration catalog is expanding. If a specific guardrail is a hard requirement, ask us — there's a good chance it's already shipped or close.
Pre-acquisition Portkey published Free (10K req/mo), Pro (usage-based), and Enterprise (custom) tiers. Post-acquisition pricing is something to watch — enterprise platform vendors often migrate self-serve plans toward enterprise contracts over 12–18 months. We don't have a confirmed change to point to today.