The short version
OpenRouter is a good product. It is also a layer between you and the model, run by a company with its own suppliers and its own terms — and in 2026 a lot of teams found out what that means at three in the morning.
This is an honest map of what else exists, what each one actually charges, and which of them are genuinely different rather than differently branded.
Not price. Price is the reason people arrive. Four things push them out.
In early 2026 OpenRouter began enforcing a clause in its terms prohibiting users from "reselling API access to AI Models or otherwise developing a competing service". Accounts started receiving 403 errors, and services built on the platform broke.
From the Hacker News thread at the time: "All Gemini and Claude models are not accessible anymore. Does anyone have any insights?" And, more bluntly: "OpenRouter just practically killed hundreds of services in a day."
OpenRouter's answer was reasonable: "We're required to comply with the terms of service of our upstream model providers… This is not a ban on using OpenRouter." Both things are true, and that is exactly the problem. You are downstream of somebody else's compliance obligations.
The comment that aged best: "It's not like you ever owned anything when you built something on top of these sorts of services… Anything relatively close to production? Fix a model version and use a provider's API."
You do not always know what you are being served. From a long-running Hacker News thread: "I love the idea of OpenRouter. I hadn't realized until recently though that you don't necessarily know what quantization a certain provider is running. And of course context size can vary widely from provider to provider for the same model." A reply added the sharper version: "some providers may apply more aggressive optimization in periods of high load."
To OpenRouter's credit, they shipped a quantization filter. But the fact that downstream tools — Roo-Code, goose, qwen-code — have all opened issues asking for ways to avoid quantized providers tells you how real the problem is.
Straight from their FAQ: "OpenRouter does not currently offer volume discounts." At $100,000/month of spend, the 5.5% credit fee is $5,500 a month with no negotiation available. That funds a lot of alternatives.
They reserve the right to expire unused credits after a year. Their COO answered the criticism candidly on Hacker News: "In practice we don't expire the credits, but have to reserve the right to, or else we have an uncapped liability literally forever." Fair — and still a reason not to prepay a large balance.
| Top-up | Fee | Effective rate |
|---|---|---|
| $5 | $0.80 | 16.0% |
| $10 | $0.80 | 8.0% |
| $20 and above | 5.5% | 5.5% |
| Crypto (USDC) | 5% | 5.0% |
| BYOK, under 1M req/month | — | 0% |
| BYOK, above 1M req/month | 5% of equivalent cost | 5.0% |
Worth stating clearly, because most "alternatives" posts get this backwards: at BYOK under a million requests a month, OpenRouter is free. If that is you, no alternative on this page will save you money.
Requesty charges a flat 5% markup on token cost, and says so plainly: a model that costs $10 per 1M from OpenAI costs $10.50 through Requesty. Refreshing, because it means no top-up fee to reverse-engineer. Free tier: 200 requests/day on free models. EU data residency on every plan.
Straight comparison to OpenRouter: 5% on tokens versus 5.5% on credits. If you top up in large amounts, OpenRouter is marginally cheaper. If you top up small, Requesty wins outright.
No token markup, and $5/month of free credits per team. BYOK carries no fee. The catch is unusual: governance features are metered per request. A provider allowlist costs $0.10 per 1,000 requests, and so does zero data retention. Custom reporting is billed separately again.
So the tokens are free and the compliance is not. For a small team on Vercel already, this is arguably the cheapest option on this page. For a regulated team, price the ZDR line before you commit.
No markup plus a 5.5% platform fee at checkout — structurally identical to OpenRouter. Rate limit is the thing to check: 7 requests per second by default, raised to 15 on request. That is low enough to matter for a busy app.
Charges 5% on credit usage, plus 1.5% for international cards, and is itself open source — so you can self-host the same software and pay nothing. 200+ models across 40+ providers, three free models at 20 requests/minute, and SOC 2 Type II. Enterprise gets volume discounts, which OpenRouter explicitly does not offer.
A different animal. These sell tokens at prices that are sometimes below the model vendor's own list. CometAPI markets a "permanent 20% discount" and says credits never expire. In our own Fable 5 comparison the spread across resellers runs from 40% below Anthropic's list price to 10% above it.
That gap is the entire reason to check rather than assume — and it is also why below-list pricing deserves a hard look. Our gateway guide walks through the three plausible explanations, one of which you should care about a great deal.
MIT-licensed, 100+ providers, 240 million+ Docker pulls. The free build includes what most managed vendors charge for: virtual keys, per-key budgets, teams, load balancing, rate limits, guardrails, and caching on Redis, Qdrant or Valkey including semantic caching.
Costs: Postgres, Redis, replicas, and someone on call. Roughly $150–400/month of infrastructure, plus the engineer nobody budgets. We work through the break-even in our OpenRouter vs LiteLLM comparison — short version, it pays above about $10k/month of model spend and rarely below.
One honest mark against it: two LiteLLM releases were briefly backdoored on PyPI in March 2026. They disclosed it clearly. Pin your versions.
Apache 2.0, written in Go, and aggressively benchmarked against LiteLLM — 0.99ms of overhead versus 40ms, 100% versus 88.78% success at 500 requests per second. Those numbers are real and they were produced by Bifrost, about a competitor's product, on hardware Bifrost chose. Read them as marketing that happens to be measurable.
Open source, self-hostable, with a managed tier at $49/month for 100,000 logs and $9 per additional 100,000. The critical detail: Portkey meters logs, not requests. Blow through your log allowance and the gateway keeps routing — it just stops recording. A team doing 500,000 requests on a 10,000-log free plan has 98% of its traffic invisible.
One more thing worth knowing before you standardise on it: Portkey was acquired by Palo Alto Networks in 2026. Roadmaps and pricing change after acquisitions.
Open source, free Hobby tier at 10,000 requests/month — but with a 10 logs per minute ingest cap, which is the real constraint, not the monthly number. Pro is $79/month. SOC 2 and HIPAA start at the $799/month Team tier.
The only two on this page that charge nothing on top of the model. Amp states it plainly: "zero markup on the providers' API pricing", $5 minimum credit purchase. OpenHands is MIT-licensed and free to self-host, with its own provider offered at cost. If your objection to gateways is philosophical rather than financial, these exist.
Cloudflare AI Gateway gives you analytics, caching and rate limiting free on every plan, with a 5% fee only if you use its unified billing. Two caveats: the free tier caps you at 100,000 logs in total, forever, and its OpenAI-compatible Unified API is currently marked deprecated in Cloudflare's own docs.
Bedrock, Azure AI Foundry and Vertex are not really alternatives to OpenRouter — they are single-vendor catalogues billed through a cloud account. They win on compliance and procurement and lose on breadth. Note that regional endpoints on Bedrock and Vertex carry a 10% premium over global ones.
| Name | Type | Open source | Fee | Free tier | Watch out for |
|---|---|---|---|---|---|
| OpenRouter (baseline) | Managed | No | 0% tokens + 5.5% on top-ups ($0.80 min) | Free models, 50–1,000 req/day | No volume discounts; credits may expire; reselling prohibited |
| Requesty | Managed | No | 5% on tokens | 200 req/day on free models | Markup applies to every token, not just top-ups |
| Vercel AI Gateway | Managed | No | 0% on tokens; ZDR and allowlist $0.10/1k requests each | $5/month credits | Governance features are metered per request |
| Eden AI | Managed | No | 0% markup + 5.5% platform fee | Pay-as-you-go | 7 req/sec default rate limit |
| LLM Gateway | Managed + OSS | Yes | 5% on credit usage + 1.5% intl card | 3 free models, 20 req/min | Smaller catalogue (200+ models) |
| CometAPI | Reseller | No | Claims ~20% below list; credits never expire | Trial credits | Below-list pricing deserves a provenance check |
| LiteLLM | Self-hosted | Yes (MIT) | $0 software; you pay infra | Entire OSS build | Postgres + Redis + on-call; no compliance certs |
| Bifrost | Self-hosted | Yes (Apache 2.0) | $0 | Full OSS | Its LiteLLM benchmarks are vendor-run |
| Portkey | Self-host + managed | Yes | OSS free; $49/mo for 100k logs | 10k logs/month | Meters logs, not requests. Acquired by Palo Alto Networks |
| Helicone | Observability gateway | Yes | Hobby free; $79/mo Pro; $799 Team | 10k req/mo | 10 logs/minute ingest cap on free |
| Amp | Managed | No | Zero markup — pure pass-through | $5 minimum purchase | Narrower product scope |
| OpenHands | Self-hosted | Yes (MIT) | $0; provider at cost | Free to self-host | Agent-oriented, not a general gateway |
| Cloudflare AI Gateway | Edge | No | Core free; 5% on unified billing | 100k logs total | Its OpenAI-compatible API is marked deprecated |
| TrueFoundry | Managed + self-host | No | $0 / $499 / $2,999 per month by request volume | 50k req/mo | Priced per request, not per token |
Every aggregator routes you to somebody else's hardware, and you are trusting them to serve the real model at full precision. This is the risk that price comparisons ignore.
Ask three questions before you commit:
| Gateway | Free tier | The real constraint |
|---|---|---|
| OpenRouter | Free models | 50 req/day, or 1,000 once you have bought $10 of credits |
| Vercel AI Gateway | $5/month credits per team | Governance features cost extra per request |
| Requesty | 200 req/day | Free models only |
| Helicone | 10,000 req/month | 10 logs per minute — this is the one that bites |
| Portkey | 10,000 logs/month | 3-day log retention |
| Cloudflare | Core features free | 100,000 logs total, forever, across all gateways |
| TrueFoundry | 50,000 req/month | 3 users |
| LiteLLM, Bifrost, OpenHands | Entirely free | You run it. That is the price. |
The API call itself is a base URL and a key — everything here speaks the OpenAI format. What does not travel:
:nitro, :floor and :exacto variants exist nowhere else.int4 providers, most alternatives give you no way to.| If… | Look at |
|---|---|
| You top up in small amounts | Requesty (5% flat) — the $0.80 minimum is what is hurting you |
| You are already on Vercel | Vercel AI Gateway — 0% on tokens, but price the ZDR line first |
| You want zero margin, on principle | Amp or OpenHands — pure pass-through |
| You spend over ~$10k/month on models | LiteLLM — the fee now funds the infrastructure |
| Your product resells model access | Anything self-hosted. OpenRouter's terms forbid it and they enforce it |
| You need SOC 2 or HIPAA today | OpenRouter (SOC 2 Type 2) or Helicone Team — not a Kubernetes cluster you just built |
| Latency is genuinely your constraint | Bifrost — but benchmark it on your own traffic, not theirs |
| You are already deep in one cloud | Bedrock / Foundry / Vertex — procurement will thank you |
| You just want the cheapest landed price | Check the table on our comparison page — the spread is 40% below list to 10% above |
It depends on why you are leaving. On price with small top-ups, Requesty's flat 5% beats OpenRouter's $0.80 minimum. On control and terms-of-service risk, self-host LiteLLM. On zero margin, Amp or OpenHands. There is no single answer, and anyone giving you one is selling something.
In early 2026 it began enforcing a terms clause prohibiting reselling API access or building a competing service, citing obligations to its upstream providers. Users reported 403 errors and services breaking with little warning.
If you top up in small amounts, yes. Requesty charges a flat 5% on tokens; OpenRouter charges 5.5% on credit purchases with an $0.80 minimum, which is 16% on a $5 top-up. On large top-ups the two are close.
Yes. Amp passes provider pricing through with zero markup, and OpenHands is MIT-licensed and free to self-host with its provider offered at cost. LiteLLM and Bifrost are free software — you pay only infrastructure.
No. Its FAQ states: "OpenRouter does not currently offer volume discounts." Enterprise customers can buy credits in bulk at discounted fees, but no percentage is published. LLM Gateway does offer enterprise volume discounts.
At high volume, a self-hosted one — LiteLLM or Bifrost cost nothing in licence. At low volume with BYOK, OpenRouter is free below 1M requests a month. The honest answer requires computing landed cost at the top-up size you actually use.
Partly. Its core features are free on all plans, but the free tier caps you at 100,000 logs in total, forever, and its OpenAI-compatible Unified API is currently marked deprecated in Cloudflare's own documentation.
Portkey charges by logged records rather than API calls. Exceed the allowance and the gateway keeps routing traffic — it just stops recording it. You go blind rather than down, which is arguably worse.
It can. Developers have reported aggregator-served models performing worse than the same model called directly, with suspicion on quantized weights or throughput-optimised endpoints. Pin the upstream provider where you can, and run your own evaluations.
If your spend is above roughly $10,000/month, or your product resells model access, or you need a VPC-only deployment — yes. Below that, the fee is usually cheaper than the engineer.
Some do. In our own Fable 5 comparison, gateway prices run from 40% below Anthropic's list to 10% above it for the identical model. Below-list pricing has plausible explanations and one worrying one — test before you trust.
The API call is a base URL and a key. What breaks is everything around it: OpenRouter's model variants, provider pinning, quantization filters, automatic fallback chains, and the sticky routing that keeps your prompt cache warm.
For real usage, LiteLLM or Bifrost — entirely free, you just run them. Among managed options, Vercel's $5/month of credits and Requesty's 200 requests/day are the most usable; Helicone's free tier is throttled to 10 logs per minute.
Possibly. Palo Alto Networks completed its acquisition of Portkey in 2026. Acquisitions routinely change roadmaps and pricing. If you are standardising on it, ask about the roadmap before you do.
OpenRouter's fee is not the reason to leave it. At BYOK under a million requests a month, it costs you nothing at all, and its routing and compliance are genuinely good.
The reason to leave is structural: you are downstream of somebody else's supplier agreements, and in 2026 a lot of teams discovered what that costs at three in the morning. If your product resells model access, or your spend is large enough that 5.5% funds an engineer, or you simply cannot afford to be 403'd — hold your own provider keys.
If none of those apply, the honest advice is boring: stay, top up in larger amounts so the $0.80 minimum stops mattering, and spend the afternoon you would have spent migrating on turning on prompt caching instead. That is worth 90% of your input bill. No gateway fee comes close.
Copyright © 2026 Best AI Gateways
Independent ranking. We may be rewarded for recommending the service we rate best and sending users to it — that reward pays for the research behind this comparison and never buys a ranking position, at no extra cost to you. Information is provided “as is” without warranty. See our Disclaimer & Terms.