The OpenAI-compatible API for every model that matters.

What "OpenAI-compatible" actually means here

AI Generate accepts the same Authorization: Bearer header you already send to OpenAI, the same {model, messages, stream, tools} JSON, and returns the same {id, choices, usage} shape. Streaming SSE chunks ship in the canonical data: {choices:[{delta:{content}}]} format that the OpenAI Python and JavaScript SDKs already consume. The only thing you change is base_url. Every model name is a string — pass claude-sonnet-4-5 and we route to Anthropic, gpt-5-codex and we route to OpenAI direct, gemini-3-pro and we route to Google, and any other model name in the catalog fans out through OpenRouter. There is no per-provider SDK to learn, no per-account billing to consolidate, no per-modality auth to manage. LangChain, llama-index, the Vercel AI SDK and any framework that already speaks OpenAI also work without modification — the wire format is what they target, and the wire format is what we serve.

Why a single Bearer token saves more than a few keystrokes

Most production AI stacks talk to three to five providers — OpenAI for code, Anthropic for analysis, Google for long context, fal or Replicate for image, Runway for video. Each is a separate billing relationship: separate invoices, separate spend caps, separate dashboards, separate Stripe webhooks if you bill customers downstream. AI Generate folds all of them behind one credit pool with one API key. Your finance team gets one line item, your operators get one dashboard, your engineers get one SDK, and your error-handling code stops branching on which provider died this morning. The 10–40% markup pays for the consolidation; the five-tier auto-discount makes the math work above $200 a month. At the strategic tier the unit price is competitive with direct-from-OpenAI billing once you account for the operational cost of running three to five separate provider relationships in parallel.

Drop-in migration with no SDK rewrite

If your codebase imports the OpenAI Python or JavaScript package, the migration is a constructor-argument change. Set base_url (Python) or baseURL (JavaScript) to https://aimarcusimage.eu/api/v1 and your AI Generate key in api_key. Every chat-completions call works without further changes — including the tool_calls block, the response_format JSON mode, the vision content[], the streaming async iterator and the standard usage object. The OpenAI client retries, timeouts and request-id headers behave the same way. For embeddings, point the same client at /v1/embeddings — five embedding models including text-embedding-3-large. For generative media the surface diverges from OpenAI (because OpenAI does not ship video or music as a single API): you POST to /v1/jobs/createTask, receive a taskId, and either poll /v1/jobs/recordInfo or wait for our webhook callback. Auth and credit pool stay shared across all surfaces.

How the volume tier ladder actually applies

Markup over upstream cost starts at 40% and drops automatically as your rolling 30-day spend climbs: 30% at $50, 22% at $200, 15% at $1,000, 10% at $5,000. The tier recomputes every six hours against the past 30 days of successful, billed calls — so a busy launch month moves you up, a quiet month does not punish you with a sudden cliff. There is no negotiation, no sales call and no contract. The current tier is visible on every /v1/me response and in your dashboard. Failed calls (upstream 5xx, internal errors, client 4xx) are excluded from the tier calculation: only the spend that would actually appear on an OpenAI invoice counts. At the strategic tier the math frequently beats per-account direct billing once you fold in the cost of running three to five separate provider relationships, each with its own SLO, its own retry budget and its own finance integration.