Skip to content
FilmOpen documentation

OpenRouter: technical notes

These notes gather what OpenRouter documents about its API, for anyone comparing platforms or building on FilmOpen. The steps for getting a key are on OpenRouter. Every page in Sources was read on 15 September 2026.

  • Console facts. What OpenRouter’s site showed on 15 September 2026, while an account was opened with Google, is marked (console).
  • “Not documented” means OpenRouter’s pages say nothing about it.

Account and money

Section titled “Account and money”
  • Sign-in: GitHub, Google or MetaMask, or an e-mail address and a password, with an optional first and last name (console). Signing up with Google led to Legal consent, accepting the Terms of Service, Privacy Policy and Model Terms, then to first-run steps asking whether the account is for an individual or an organisation (console).
  • Who can sign up: people aged 18 or over (terms, §2). OpenRouter publishes no list of countries, but some model providers don’t let people in certain countries or regions use their models (terms, §5.7).
  • Phone, identity checks and two-factor sign-in: not documented. Signing up with Google asked for no phone or identity check (console).
  • Workspaces: a new account’s first workspace, made at sign-up, keeps its API keys, usage tracking and billing together (console).
  • Billing: prepaid credits in US dollars, from $5 to $25,000 a purchase. Unused credits can be refunded if asked within 24 hours of a purchase; fees and crypto payments are never refunded (terms, §4.1). OpenRouter may expire credits 365 days after purchase (terms, §4.2).
  • Fees: 5.5% on card purchases, at least $0.80, and 5% in crypto (FAQ).
  • Payment methods: cards, AliPay and USDC (FAQ). A first-run step asks for a card number and an amount (console).
  • Tax: not documented.
  • Auto-recharge: optional. When buying credits, a user can have OpenRouter charge the chosen payment method whenever credits fall below a threshold the user sets, which the terms call Auto Recharge (§4.2). Whether it is on when first offered: not documented.
  • Limits:
    • Per key: a spending limit in US dollars, which can reset daily, weekly or monthly (creating a key).
    • When requests stop: a request gets 402 once the account’s balance is below zero, or a key’s remaining limit reaches zero (limits).
    • Guardrails add budgets and rules on models and providers, for keys, members or workspaces (guardrails).
    • A cap on the whole account: not documented.
  • Usage: OpenRouter’s Activity page lists requests by model, provider and key (FAQ).

Keys

Section titled “Keys”
  • Where: the API keys page, where a key gets a name and, optionally, a credit limit (authentication).
  • Made at sign-up: yes. The first-run step Your workspace is ready shows the first workspace’s API key, saying it is the only time the full key will be shown (console).
    • In its field the key is shortened.
    • In its Make your first request code sample the key is whole, so a screen capture of that step captures the key (console).
  • What you copy: the key, which starts sk-or-v1- (console).
  • Kinds: API keys, which call the models, and management keys, which create and manage API keys but can’t call the completion endpoints (management keys).
  • Keys created by API need a management key (creating a key).
    • The request: name, limit in US dollars, limit_reset (daily, weekly, monthly or none), include_byok_in_limit, and expires_at (an ISO 8601 time in UTC, with seconds).
    • The key itself is returned only in that response.
  • Checking a key: GET https://openrouter.ai/api/v1/key returns the key’s label, limit, remaining limit, usage and expiry (current key), and OpenRouter’s limits page names it as the way to check what a key has left (limits). Whether it costs anything: not documented.
  • The balance by API: GET /api/v1/credits needs a management key (credits).
  • Leaks: OpenRouter uses GitHub secret scanning to find its keys in public code, and e-mails the owner (authentication).
  • Not documented: scopes for API keys.

Calling the API

Section titled “Calling the API”
  • Base URL and headers: https://openrouter.ai/api/v1, with Authorization: Bearer <key>. Optional HTTP-Referer and title headers credit an app in OpenRouter’s rankings (authentication, API overview).
  • Text: POST /api/v1/chat/completions, in the shape of OpenAI’s Chat Completions. Parameters a model doesn’t support are ignored (API overview). A stateless Responses API is at /api/v1/responses (Responses API).
  • Streaming (streaming):
    • The stream: server-sent events. Comment lines such as : OPENROUTER PROCESSING can be ignored, and usage comes in the last chunk.
    • An error after the stream has started arrives as an event with finish_reason: "error".
    • Cancelling a stream stops processing and billing only with providers that support it, such as OpenAI and Anthropic; not with others, such as Google and Amazon Bedrock.
  • Routing:
    • Choosing a provider. OpenRouter picks one for each request, weighing price and skipping providers with recent outages. The provider object can order, allow or ignore providers (provider selection).
    • Fallbacks. A models list falls back to other models; the call is billed at the model that answered (model fallbacks).
  • Rate limits (limits):
    • Free models allow 20 requests a minute and 50 a day, or 1,000 a day once $10 of credits has been bought.
    • Capacity is governed globally, so more accounts or keys don’t raise the limits.
  • Errors (errors):
    • 400: bad parameters;
    • 401: an invalid or disabled key;
    • 402: no credits left;
    • 403: moderation or a guardrail;
    • 408: a timeout;
    • 429: rate limited;
    • 502: the model is down or answered badly;
    • 503: no provider meets the request’s routing requirements.
  • Idempotency keys: not documented.
  • SDKs: TypeScript, Python and Go. There is none for Dart, so FilmOpen calls the HTTP API (SDKs).

The models FilmOpen uses

Section titled “The models FilmOpen uses”

Prices per million tokens, from OpenRouter’s public endpoints API, one address per model, such as Claude Opus 5’s:

Model Input / output Cache read / write Longer prompts Context, tokens
Claude Opus 5 $5 / $25 $0.50 / $6.25 — 1,000,000
Claude Sonnet 5 $2 / $10 $0.20 / $2.50 — 1,000,000
GPT-5.6 Sol $2 / $10, marked 50% off $0.20 / $2.50 272,000 or more: $4 / $15 1,050,000
GPT-5.6 Terra $2 / $12 $0.20 / $2.50 272,000 or more: $4 / $18 1,050,000
Gemini 3.1 Pro $2 / $12 $0.20 / $0.375 Above 200,000: $4 / $18 1,048,576
Gemini 3.8 Flash $0.75 / $3.75, marked 50% off $0.075 / about $0.04 — 1,048,576
  • Providers: the table gives each model’s maker’s own endpoint (endpoints API).
    • Anthropic’s models are also served by Claude Platform on AWS, Amazon Bedrock, Azure and Google Vertex; regional endpoints charge 10% more.
    • OpenAI’s models are also served by Azure and Amazon Bedrock. GPT-5.6 Sol’s discount is on OpenAI’s own endpoints only, so Sol costs more there: $4.40 / $22 on Bedrock, $5 / $30 on Azure and $5.50 / $33 on Azure’s US and EU endpoints. GPT-5.6 Terra costs OpenAI’s own $2 / $12 on Azure, and 10% more on Azure’s US and EU endpoints and on Bedrock.
    • Gemini is served by Google AI Studio and Vertex, each with a cheaper flex endpoint and a dearer priority one.
  • Discounts: OpenRouter’s discounted models list GPT-5.6 Sol and Gemini 3.8 Flash at 50% off, with no end date.
  • Prompt caching: writing to Anthropic’s cache costs 1.25 times the input price for five minutes, or twice for an hour, and reading costs a tenth (prompt caching).
  • Images: the catalogue’s Nano Banana models are on OpenRouter too, made through POST /api/v1/images (images API). FilmOpen’s first version makes images through fal.

Media in

Section titled “Media in”
  • Images go in a message as an image_url part: an https URL or a base64 data URL, in PNG, JPEG, WebP or GIF (image understanding). Size limits: not documented.
  • PDFs go as a file part, by URL or as base64 (PDFs).
  • Audio goes as base64 in an input_audio part; a URL is not accepted (audio).
  • The images API’s reference images: up to 16, as base64 or http(s) URLs; a request that is too large gets 413 (images API).
  • What a URL given as input must satisfy: not documented.

Media out

Section titled “Media out”
  • Images from /api/v1/images come back as base64 with their media type (images API), and a failed generation is not billed (image generation).
  • What OpenRouter keeps:
    • Logging: prompts and completions are not logged by default (FAQ), and a user can opt in (data collection).
    • Training: OpenRouter does not use inputs or outputs to train models (privacy).
    • Zero data retention: requests can be limited to providers that keep nothing (zero data retention).
  • Writing results straight to your own storage: not documented.

Usage and cost

Section titled “Usage and cost”
  • Every response reports usage, including its cost in credits, without being asked (usage accounting).
  • Per key: GET /api/v1/key reports a key’s usage, including daily, weekly and monthly totals (current key).
  • Requests that fail or are cancelled:
    • Prompts. A provider may still charge for processing a prompt that produced nothing (errors).
    • Fallbacks. Only the model that answered is billed (model fallbacks).
    • Cancelled streams. Cancelling stops billing only with providers that support it (streaming).

Open questions

Section titled “Open questions”
  • When does the 50% discount on GPT-5.6 Sol and Gemini 3.8 Flash end?
  • Does checking a key with GET /api/v1/key cost anything?
  • Is Auto Recharge on or off when it is first offered, and what does it ask for?
  • Can the API keys page set a limit’s reset period and an expiry, as the API can?
  • Does OpenRouter add tax to a purchase of credits?
  • What do its files and logging pages add to how media is taken and kept? They were not read.

Sources

Section titled “Sources”

Read on 15 September 2026: