Resposta direta
Esta página explica como os times usam o gateway compatível com OpenAI do NextModel. Migration, billing boundaries, limits, and the questions support answers most. Ela adiciona os passos práticos, as notas de configuração e as perguntas frequentes.
Do I need to change code to migrate from OpenAI or Anthropic?
No. Swap two values: base_url → https://api.nextmodel.app (OpenAI tools usually append /v1 themselves; Anthropic tools take the bare host because the gateway serves /v1/messages on the same path) and the key → a NextModel key. SDKs, request shapes, and streaming work unchanged. The one field to also set is model — it must be a catalog id from GET /v1/models.
What exactly is a failed request billed?
Nothing. Reservations settle only after a response is delivered in full; provider errors, mid-stream failures, and client disconnects all roll back. Every response's X-NextModel-Billing-Status header proves which happened.
Why does my model call return model_not_supported?
The id is not in the live catalog. Copy it from GET /v1/models including the provider prefix; display names and bare shortcuts are not valid ids. See Model ID rules.
Do you support Gemini- or DashScope-native protocols?
Not yet. Gemini, Qwen and other vendor-native families are called through the OpenAI-compatible endpoints with their catalog ids; dedicated native paths are on the roadmap and currently return 501 protocol_runtime_not_routed.
Is streaming supported?
Yes — set stream: true for SSE on chat, responses, and messages endpoints.
Is function calling / tool use supported?
Yes on models whose catalog capabilities include Function Calling (visible in GET /v1/models). Pass tools in the standard OpenAI or Anthropic shape for the endpoint you are calling.
What are the rate limits?
Rate limits are per-account and per-key; exceeding them returns 429 with a Retry-After header. The current deployment allows 6,000 requests per minute per IP and per account. If you need a higher ceiling, contact support with your expected QPS.
Can I create multiple API keys?
Yes, and you should: split by environment (dev/staging/prod) or by project so /v1/activity and the dashboard attribute usage per key. Disabling a leaked key stops its calls immediately; disabled keys cannot incur charges.
How do I verify I got a real model response?
Beyond the response's model field and billing headers, NextModel issues verifiable provenance receipts — see /verifiable and the AI provenance doc — so an auditor can check the receipt against the gateway's proof, not just take the provider's word.
My stream died halfway. What do I check?
First confirm a non-streaming request works, then check client timeouts and proxies. A stream that does not complete cleanly rolls back the reservation — check X-NextModel-Billing-Status on the failed request — and the codes upstream_stream_idle_timeout / upstream_stream_total_timeout on the error identify where it stopped.
Does the platform modify my prompts or responses?
No. The gateway forwards requests and relays provider responses; it does not rewrite contents. Failed egress translation is surfaced as a 502 egress_translation_failed rather than silently mangled output.
Is my data recorded?
The gateway logs metadata — time, model, token counts, cost, status — for billing and reliability. Prompt and completion bodies are not retained beyond request logs needed for the product's dashboard and idempotent-replay guarantees.