OpenRouter: fees, free model limits and the Stripe deal

This page is built from OpenRouter's own pricing page, docs, terms, public model list and blog, the makers' price pages, independent tests by Moonshot AI, Mo Moustafa and the GateScope researchers, a competitor's latency test, and user reports on Hacker News, Trustpilot and Product Hunt, all checked on October 21, 2026. OpenRouter's claims are labelled as its own.

According to AICPB's traffic rankings, openrouter.ai had 21.3 million visits in September 2026, down 4.7% on August: 1st among AI model routers.

What OpenRouter sells

OpenRouter sells one API key and one prepaid balance for other companies' models. Its public model list returned 686 model IDs on October 21, 2026, including 63 image-output and 30 video-output models, from 94 model makers and 113 hosting providers; its pricing page says 500+ models from 80+ providers. Popular open-weight models have many competing hosts: DeepSeek V4.1 Flash had 37 endpoints that day.

By default OpenRouter spreads requests across a model's hosts, favouring cheaper ones, and falls back automatically. Per request you can pin providers, sort by price, throughput or latency, filter by quantization, cap the price or require zero data retention (routing docs). Since March 2026 its Auto Exacto routing moves tool-calling requests away from hosts with poor tool-call results.

Stripe agreed to acquire OpenRouter, Inc. of New York on August 19, 2026. OpenRouter said its name and integrations would stay and expected to close "in the coming weeks"; no closing had been announced by October 21, 2026, and nothing has been said about fees.

Prices against the official APIs

OpenRouter's FAQ says there is no markup on inference pricing. It charges 5.5% when you buy credits on the Standard plan and 8% on Business. Bring-your-own-key traffic is free up to 25,000 USD a month at list price, then 5%.

OpenRouter prices come from its model list and endpoint lists. For closed models OpenRouter costs 5.5% more than the maker once the fee is counted. Savings come on open-weight models, where hosts compete; the cheapest DeepSeek V4.1 Flash endpoint runs a 4-bit (fp4) build. OpenRouter's June 2025 fee post also set a 0.80 USD minimum fee on card purchases, which would be 16% of a 5 USD top-up; the live pricing page no longer states it.

What public tests show

• Moonshot AI's K2 Vendor Verifier (November 2025) sent 4,000 tool-calling requests to each host. Hosts reached through OpenRouter scored 83.05% (Chutes) to 96.46% (NovitaAI) tool-call schema accuracy on Kimi K2-thinking, against 100% on Moonshot's API. It measures the hosts, not OpenRouter's own layer.
• Mo Moustafa (September 2026), who runs an iMessage assistant with about 18 million messages, measured DeepSeek V4 Flash at about 90% on GPQA Diamond and 81% on TAU-Bench on the first-party host, against about 75% and 58% on DigitalOcean. One host carried about 20% of his traffic but 92% of the empty replies.
• GateScope (ACM IMC '26 paper, April 2026) audited 10 anonymised gateways. The one coded "o*er" matched the claimed model in 86.91% to 97.09% of responses against 96.00% to 98.18% for the official APIs, with no billing gap on gpt-4o; other gateways fell below 60% or overbilled by up to 62.8%. The paper does not name OpenRouter.
• Max Woolf (May 2026) showed that OpenRouter's public rankings measure token volume: Tencent's Hy3 preview topped them through a single provider, with about 98% input tokens. Read the rankings as usage, not quality.
• china-llm.com (July 2026), three runs from one location, measured 2.54 seconds to first token through a routed DeepSeek path against 0.71 seconds direct. The site advertises a competing aggregator.
• Opper, a competing router (April 2026), timed 200 GPT-4.1 calls: OpenRouter reached the first token in 0.640 seconds against 0.712 seconds for OpenAI direct, with about 10% lower throughput.
• OpenRouter's own figures (March 2026) say Auto Exacto cut GLM-5 tool-call errors from about 8% to about 1%.

What users report

Hacker News threads praise the single key, spending caps and quick model switching; in the Series B thread Simon Willison called it the lowest-friction way to try every model. Recurring complaints: results that vary by host (September 2026), a fee some users learned about late and account-level region blocks on OpenAI, Anthropic and Google models (April 2026), Gemini and Claude access disappearing from accounts (March 2026), and a stolen key that spent 100 USD despite a 10 USD daily limit. Trustpilot shows 1.6 out of 5 from 67 reviews, 84% one star, about unexpected charges, vanished credits, refused refunds, bans and slow support. Product Hunt shows 5.0 from 54 reviews, which it notes come almost entirely from founders. The Stripe deal thread mixed praise with doubts about the router's neutrality under a payments company.

MCP, Agent Skills and coding agents

• MCP server: OpenRouter hosts an official remote MCP server at https://mcp.openrouter.ai/mcp. An OAuth sign-in mints a dedicated key that expires after 7 days with a 10 USD default spend cap. Its tools search models, prices, benchmarks and docs, read your credits and generation costs, and send billable test messages or images. OpenRouter says to call the API directly in production.
• MCP from the API: the Responses API schema includes an OpenAI-style mcp tool type with a server URL; the docs do not say which models it works with. A cookbook shows converting MCP tools to OpenAI tool definitions, and the @openrouter/mcp package connects remote MCP servers to its Agent SDK.
• Agent Skills: the official OpenRouterTeam/skills repository held 18 SKILL.md skills on October 21, 2026, covering its SDKs, analytics, models, images, video, speech and OAuth, installable as a Claude Code plugin or with gh skill install. Its Anthropic Messages reference does not mention Anthropic's Skills API.
• Coding agents: guides cover Claude Code, Claude Desktop, Codex CLI and the Codex desktop app, Cursor, Hermes Agent, Junie CLI, OpenClaw and OpenCode. Claude Code uses the Anthropic-compatible endpoint, which OpenRouter only guarantees with Anthropic as the first-choice provider (guide).
• API formats and SDKs: Chat Completions, Responses and Anthropic Messages, a Batch API, structured outputs, prompt caching and server tools such as web search; no Gemini-native endpoint. Official TypeScript, Python and Go SDKs, Vercel AI SDK and LangChain guides, and an llms.txt index.

Privacy, terms and who runs it

OpenRouter logs request metadata but not prompts or completions unless you opt in, which earns a 1% discount. You can block providers that may train on data and enforce zero data retention per account or per request. Its trust page lists SOC 2 Type 2 behind a request form. Processing inside the EU or US is limited to Business and Enterprise.

Its terms set a 5 USD minimum and 25,000 USD maximum per purchase, allow refunds only within 24 hours, never refund the platform fee or crypto payments, let unused credits expire 365 days after purchase, and forbid reselling API access. OpenRouter published a post-mortem for outages of 38 and 35 minutes in February 2026. It raised a 113 million USD Series B led by CapitalG in May 2026 (TechCrunch).

How it compares with Requesty, Vercel and Cloudflare

• Requesty also charges list price, adds a flat 5% to top-ups, has an EU endpoint in Frankfurt and a free plan of 200 requests a day; it lists no video models.
• Vercel AI Gateway passes list prices through with no platform fee.
• Cloudflare AI Gateway charges 5% on credit purchases and nothing for routing your own keys.
• Kilo Gateway has run largely on OpenRouter and adds 5% to purchases.
• NanoGPT charges no deposit fee and takes cryptocurrency.

Who should skip it

Skip OpenRouter if you need identical model behaviour without pinning hosts yourself, plan to resell API access, live where the large labs block access, or cannot accept a 24-hour refund window. For production on open-weight models, pin hosts you have checked and keep fallbacks on.