Portkey
The best-instrumented AI gateway you can buy, and since 29/05/2026 it is a Palo Alto Networks product — so audit the roadmap, not just the feature list. The MIT gateway on GitHub is the part nobody can take away from you.
Element scores
Strengths
Portkey does one job and does it better than anyone: it sits in the path of every model call and every tool call, and tells you what happened. OpenTelemetry tracing, 21+ metrics, per-request cost attribution, budgets per key and per team, guardrails that actually block in the request path, and prompt versioning — in one product, reachable in three lines of code. The MCP Gateway is the standout: a registry with OAuth injection, per-tool permissions, identity forwarding and a logged call for every tool invocation, which is the thing most teams discover they need about six weeks into running agents. The numbers behind it are real and officially stated: 500B tokens a day, 120M requests, $180M in annualised AI spend under management across 24,000+ organisations. And the core gateway is MIT-licensed with 12,700+ GitHub stars — you can npx it, self-host it, and never send a byte to a vendor.
Honest dings
The ownership changed. Palo Alto Networks closed the acquisition on 29/05/2026 and Portkey is now the engine inside Prisma AIRS AI Gateway, generally available since July 2026. The pricing page's own upgrade button sends you to a Palo Alto form. Nothing has been taken away yet and the open-source repo is still MIT — but a startup control plane and an enterprise security platform have different roadmaps, and an SMB buying today should assume the product will be pulled toward large-enterprise security. Beyond that: the log-metered pricing bites hardest exactly when agents make it valuable, guardrails skip image inputs and only check the last message, evaluation stops at a feedback API and an Enterprise template pack, and autonomous fine-tuning is still private beta. It is a gateway, so retrieval and vector storage are not its job and score accordingly.
Sources (15) — every claim traceable
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- paloaltonetworks.com/company/press/2026/palo-alto-n… — Official press release: acquisition closed 29/05/2026; Portkey becomes the core AI Gateway for Prisma AIRS (accessed 2026-08-25)
- investors.paloaltonetworks.com/news-releases/news-r… — Intent-to-acquire release 30/04/2026: trillions of tokens per month, 99.99% uptime claim, agent identity security positioning (accessed 2026-08-25)
- portkey.ai/blog/series-a-funding — Official metrics table (19/02/2026): 24,000+ organisations, 500B+ daily tokens, 120M daily requests, $180M+ annualised spend under management, 60+ providers, 1,600+ models; $15M Series A led by Elevation Capital with Lightspeed; basic enterprise gateway made free (accessed 2026-08-25)
- portkey.ai/pricing — Live pricing verified: Developer free (10k logs, 3-day retention), Production $49/month (100k logs, +$9/100k), Enterprise custom; upgrade button now routes to Palo Alto Networks (accessed 2026-08-25)
- portkey.ai/docs/product/guardrails — Official guardrails docs: 20+ deterministic checks, LLM-based checks (prompt injection, gibberish), partner integrations (Aporia, SydeLabs, Pillar), 7 action types, plan gating by tier (accessed 2026-08-25)
- docs.portkey.ai/docs/product/guardrails/capabilities — Endpoint coverage table: guardrails on chat/completions, embeddings (input only), messages, responses; no image inputs; only the last message evaluated (accessed 2026-08-25)
- portkey.ai/docs/product/observability — Official observability docs: OpenTelemetry compliance, 21+ metrics, traces, filters, custom metadata, feedback, budget limits (accessed 2026-08-25)
- portkey.ai/docs/product/mcp-gateway — Official MCP Gateway docs: server registry, OAuth/API-key injection, workspace and per-tool access control, identity forwarding, rate limits, guardrails and approval workflows, HTTP Streamable transport (accessed 2026-08-25)
- portkey.ai/docs/product/prompt-engineering-studio — Official prompt docs: playground across 1,600+ models, side-by-side comparison, templates, partials, versioning with rollback, prompt library and API (accessed 2026-08-25)
- portkey.ai/docs/product/autonomous-fine-tuning — Official: autonomous fine-tuning is in private beta; log-derived datasets, 10+ provider targets (accessed 2026-08-25)
- github.com/Portkey-AI/gateway — Open-source repo: MIT licence, 12,700+ stars, 1,250+ forks, created 08/2023; Gateway 2.0 pre-release merging the enterprise core into open source (accessed 2026-08-25)
- braintrust.dev/articles/ai-gateway-comparison-2026 — Independent gateway comparison: ranks Portkey second, notes teams needing evaluation-driven release gating still require a separate eval workflow (accessed 2026-08-25)
- spheron.network/blog/ai-gateway-litellm-portkey-kon… — Independent comparison flagging that semantic caching, guardrails and prompt versioning are primarily hosted-SaaS features and that guardrails create ecosystem lock-in (accessed 2026-08-25)
- truefoundry.com/blog/portkey-pricing-guide — Independent pricing analysis of the recorded-logs metric: cost scales with request volume and retention, separate from provider model spend (accessed 2026-08-25)
- tracxn.com/d/companies/portkey/__ZBFkMQ22qjERQNfNQH… — Company profile: founded 2023 in San Francisco by Rohit Agarwal and Ayush Garg, $18.1M raised, acquired by Palo Alto Networks (accessed 2026-08-25)