The best Portkey alternative depends on which part of Portkey you used. For a free self-hosted gateway, LiteLLM (MIT) and Bifrost (Apache 2.0, Go, 11 microseconds of overhead at 5,000 RPS) are the direct replacements. For one key and one wallet with no infrastructure, OpenRouter or Vercel AI Gateway. For observability first, Helicone. For enterprise Kubernetes with governance, TrueFoundry or Kong. For AI coding organisations whose spend lives on laptops rather than in a proxy, Continuum. Portkey's own MIT-licensed gateway is also a legitimate alternative to hosted Portkey and preserves your routing config.
- There is no one-to-one replacement. Portkey is four products; pick by the one you needed.
- Self-hosted and free: LiteLLM (MIT) or Bifrost (Apache 2.0, Go).
- Hosted wallet, no infra: OpenRouter (5.5% on top-ups) or Vercel AI Gateway (zero markup).
- Observability first: Helicone, $79/month Pro, $799/month Team.
- Enterprise Kubernetes: TrueFoundry ($499/month Pro) or Kong ($100/month per LLM model on Plus).
- Do not overlook self-hosting Portkey's own MIT gateway: it keeps your routing and guardrail config.
Why people are looking right now
Palo Alto Networks completed its acquisition of Portkey on 29 May 2026 (terms not disclosed), and shipped Prisma AIRS AI Gateway to general availability on 16 July 2026. The portkey.ai homepage now leads with "Portkey is now PRISMA AIRS AI Gateway".
Neither the press release nor the GA announcement says anything about what happens to the self-serve Developer and Production tiers. That silence is the reason for the search volume. It is not evidence of anything being shut down; it is evidence that self-serve is not the story the acquirer is telling.
The options at a glance
| Option | Shape | Price | Best at |
|---|---|---|---|
| LiteLLM | Self-hosted proxy, MIT | Free; Enterprise custom, sized to request capacity | 100+ providers, virtual keys, budgets, the default self-host |
| Bifrost | Self-hosted Go binary, Apache 2.0 | Free | Raw speed: 11 microseconds overhead at 5,000 RPS |
| OpenRouter | Hosted aggregator | List price; 5.5% on Stripe top-ups | Widest catalogue, one key, one wallet |
| Vercel AI Gateway | Hosted aggregator | List price, zero markup | Four API shapes, coding-agent endpoints |
| Helicone | Proxy plus observability | $0 / $79 / $799 per month | Logs, traces, and analytics as the primary product |
| TrueFoundry | Enterprise platform | $0 / $499 / $2,999 per month | VPC, on-prem, air-gapped Kubernetes |
| Kong AI Gateway | Plugins on Kong Gateway | $100/month per LLM model on Plus | You already run Kong; MCP and A2A governance |
| Cloudflare AI Gateway | Proxy over your own keys | Free core; 5% on Unified Billing credits | Caching and rate limiting on provider accounts you keep |
| Continuum | Coding-agent workbench | Free; hosted inference optional | Agent spend on laptops that no proxy can see |
By job
You wanted a free gateway you control
LiteLLM is the default answer and the one most teams land on. MIT-licensed, 100-plus providers behind an OpenAI-compatible API, with virtual keys, spend tracking, budgets, rate limits, fallbacks, logging, and Prometheus metrics in the free tier. Enterprise adds SSO and SCIM, OIDC/JWT auth, audit logs, secret managers with key rotation, org and team admin, a multi-region control plane, and 24/7 support with SLAs. LiteLLM prices Enterprise to your annual gateway request capacity and deployment architecture, explicitly not per token, and does not publish a figure.
Bifrost is the answer if throughput is the binding constraint. Apache 2.0, written in Go, from Maxim AI, 7.4k stars, and benchmarked at roughly 11 microseconds of overhead per request at 5,000 RPS on a t3.xlarge with a 100% success rate. It covers 23-plus providers, adaptive load balancing, cluster mode, guardrails, semantic caching, MCP, and budget governance. Note it is a Maxim project and has nothing to do with TrueFoundry, despite the two names appearing together in a lot of comparison content.
You wanted one key and one bill, with no servers
OpenRouter has the widest catalogue and passes inference through at provider list price. The cost is the load fee: 5.5% with an $0.80 minimum on Stripe credit purchases, 5% on crypto. BYOK is free up to $25,000 a month of equivalent spend on pay-as-you-go and $200,000 on Enterprise, then 5%.
Vercel AI Gateway charges zero markup and no platform fee on tokens, including with BYOK, and you only pay payment processing on top-ups. It speaks four API shapes (AI SDK, OpenAI Chat Completions, OpenAI Responses, Anthropic Messages), which matters if your clients are split between Codex and Claude Code. The catch is that BYOK requires purchased credits and BYOK spend cannot be capped by a budget.
You wanted the dashboards, not the routing
Helicone is observability-led with a gateway attached, which is the inverse of Portkey. Hobby is free with 10,000 requests and 1 GB of storage; Pro is $79 a month with unlimited seats, alerts and reports, and its HQL query language; Team is $799 a month with five organisations plus SOC 2 and HIPAA; Enterprise adds SAML SSO, a custom MSA, and on-prem. If what you actually valued in Portkey was the log and trace view, this is the cleaner buy.
You wanted enterprise governance in your own cluster
TrueFoundry is a full LLMOps platform with the gateway as one layer: RBAC, guardrails, semantic caching, MCP, and deployment into VPC, on-prem, air-gapped, or multi-cloud Kubernetes. Developer is free at 50,000 requests a month for 3 users, Pro is $499 a month at 1 million requests for 10 users, Pro Plus is $2,999 a month for 25 users, and Enterprise is custom from 10 million requests. A self-hosted gateway plane runs roughly $600 to $1,000 a month.
Kong AI Gateway is the right answer only if Kong is already your API gateway, in which case it is close to free marginal effort. It is also the only option here that governs MCP and agent-to-agent traffic with dedicated plugins. Konnect Plus prices AI Gateway at $100 a month per unique LLM model with five included, which is a bad shape for wide model catalogues.
Your real problem is AI coding agents
This is a different problem wearing gateway clothing. If the spend you cannot explain comes from Claude Code, Codex, Cursor, and Gemini CLI on developer machines, no proxy will find it: those tools talk directly to the provider from a laptop, often on a flat-rate subscription that a gateway would convert into metered tokens. Continuum reads the local session files they already write and attributes cost by repo, model, day, and person, then adds weekly spend caps, model policy, and member approvals at the org level.
What migration actually costs
The base-URL swap is the easy part and takes an afternoon. What does not carry over is the reason migrations slip.
| Thing | Carries over? |
|---|---|
| Client code (OpenAI-compatible) | Yes. Change the base URL and the auth header |
| Model routing and fallback order | Usually. Every gateway expresses it differently |
| Virtual keys and budget limits | No. Rebuild in the new tool's model |
| Prompt templates and versions | No. Export them before you cancel |
| Guardrail configurations | No. Rules are vendor-specific |
| Historical logs and traces | No. Production retention is 30 days anyway |
| Cost attribution history | No |
Five things to check before you commit
- Which API shapes do your clients need? Anthropic Messages and OpenAI Responses are not universally supported.
- Does the meter match your traffic shape? Log-count pricing punishes cheap high-volume calls; per-model pricing punishes wide catalogues.
- What is the log retention, and does your audit process need longer?
- Can you cap spend, and does the cap cover BYOK? On Vercel it does not.
- What happens if the vendor is acquired next year? An OSS licence is the only real answer to that question.
Questions people ask
What is the best alternative to Portkey?
It depends which part of Portkey you used. LiteLLM is the default free self-hosted gateway. Bifrost is the fastest self-hosted option. OpenRouter and Vercel AI Gateway are the hosted one-key-one-wallet choices. Helicone is the pick if observability was the real product. TrueFoundry and Kong cover enterprise Kubernetes. And self-hosting Portkey's own MIT gateway is the cheapest move of all if you only needed routing.
Is there a free open-source Portkey alternative?
Several. LiteLLM is MIT-licensed with 100-plus providers, virtual keys, budgets, rate limits, and fallbacks free to self-host. Bifrost is Apache 2.0, written in Go, with about 11 microseconds of overhead at 5,000 RPS. Portkey's own gateway is also MIT and can be self-hosted, keeping your routing and guardrails while dropping the hosted-only prompt management, semantic caching, and dashboards.
Portkey vs LiteLLM: which should I use?
LiteLLM if you want to run it yourself, pay nothing, and treat the gateway as infrastructure: 100-plus providers, virtual keys, budgets, and Prometheus metrics, with Enterprise adding SSO, SCIM, and audit logs at custom pricing. Portkey if you want guardrails, prompt management, and observability as one hosted product and are happy paying $49 a month plus log overages for it. LiteLLM has the better story on control; Portkey has the better story on breadth in a single tool.
Why are people looking for Portkey alternatives in 2026?
Palo Alto Networks completed its acquisition of Portkey on 29 May 2026 (terms not disclosed) and rebranded the product as Prisma AIRS AI Gateway, generally available from 16 July 2026. Neither announcement addressed the future of the self-serve Developer and Production tiers, which is what is driving the search. The technology is unchanged and the MIT gateway is still MIT.
Is Bifrost made by TrueFoundry?
No. Bifrost is built and maintained by Maxim AI (maximhq on GitHub), Apache 2.0, written in Go, with enterprise deployments offered through getmaxim.ai. TrueFoundry is a separate company with its own commercial AI Gateway. The two names appear together frequently because TrueFoundry publishes comparison content about Bifrost, not because they are the same product.
What if my problem is coding agents rather than an application?
Then a gateway is probably the wrong tool. Claude Code, Codex, Cursor, and Gemini CLI run on developer laptops and talk directly to providers, usually on flat-rate subscriptions that a metered gateway would make more expensive. Continuum reads the session files those tools already write and gives per-repo, per-model cost plus org-level weekly spend caps, model policy, and member approvals, without moving anyone off their subscription.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.
- Portkey pricing tiers and the Prisma AIRS banner
- Palo Alto Networks completes acquisition of Portkey 29 May 2026 completion; terms not disclosed
- LiteLLM pricing free self-hosted tier, Enterprise sized to request capacity
- maximhq/bifrost on GitHub Apache 2.0, Go, 11 microseconds at 5,000 RPS, Maxim ownership
- Helicone pricing Hobby, Pro $79, Team $799 tiers
- TrueFoundry pricing Developer, Pro $499, Pro Plus $2,999 tiers
- OpenRouter FAQ 5.5% credit fee and BYOK allowances
- Kong Konnect pricing per-LLM-model AI Gateway pricing