Models & routing
Provider keys
Use your OpenAI, Anthropic or DeepSeek keys. One key per provider is free; add fallbacks for 3%.
Pricing by providerThe enabled-key count is frozen when each request is admitted. Temporary health or balance state does not change the fee.
- One enabled key0% MiniRouter service fee for that provider.
- Multiple enabled keys3% of that provider's published-rate usage, including a first-key success.
Your provider bills inference separately. Private discounts, provider credits and taxes do not reduce the 3% reference fee.
How the fee works across providers
- One OpenAI key
- 0% for OpenAI.
- One OpenAI + one Anthropic key
- 0% for each provider.
- Two OpenAI + one Anthropic key
- 3% on OpenAI usage; 0% on Anthropic usage.
- One enabled + one disabled backup
- 0%. Disabled keys do not count.
Qualified requests
Text generation and OpenAI embeddings, sent directly to your provider. Copy a model ID and use its supported endpoint.
46 models available with your keys
OpenAI
- GPT-6 AstraChat Completions · Responses
openai/gpt-6-astra - GPT-6 SolChat Completions · Responses
openai/gpt-6-sol - GPT-6 LunaChat Completions · Responses
openai/gpt-6-luna - GPT-5.6 SolChat Completions · Responses
openai/gpt-5.6-sol - GPT-5.6 TerraChat Completions · Responses
openai/gpt-5.6-terra - GPT-5.6 LunaChat Completions · Responses
openai/gpt-5.6-luna
More OpenAI models 27
- GPT-5.5Chat Completions · Responses
openai/gpt-5.5 - GPT-5.5 ProResponses
openai/gpt-5.5-pro - GPT-5.4Chat Completions · Responses
openai/gpt-5.4 - GPT-5.4 MiniChat Completions · Responses
openai/gpt-5.4-mini - GPT-5.4 NanoChat Completions · Responses
openai/gpt-5.4-nano - GPT-5.4 ProResponses
openai/gpt-5.4-pro - GPT-5.2Chat Completions · Responses
openai/gpt-5.2 - GPT-5.3 CodexResponses
openai/gpt-5.3-codex - GPT-5.2 CodexResponses
openai/gpt-5.2-codex - GPT-5.2 ProResponses
openai/gpt-5.2-pro - GPT-5.1 ThinkingChat Completions · Responses
openai/gpt-5.1-thinking - GPT-5Chat Completions · Responses
openai/gpt-5 - GPT-5 MiniChat Completions · Responses
openai/gpt-5-mini - GPT-5 NanoChat Completions · Responses
openai/gpt-5-nano - GPT-5 ProResponses
openai/gpt-5-pro - o3Chat Completions · Responses
openai/o3 - o4-miniChat Completions · Responses
openai/o4-mini - o3-proResponses
openai/o3-pro - GPT-4.1Chat Completions · Responses
openai/gpt-4.1 - GPT-4.1 MiniChat Completions · Responses
openai/gpt-4.1-mini - GPT-4.1 NanoChat Completions · Responses
openai/gpt-4.1-nano - GPT-4oChat Completions · Responses
openai/gpt-4o - GPT-4o MiniChat Completions · Responses
openai/gpt-4o-mini - GPT-4 TurboChat Completions · Responses
openai/gpt-4-turbo - Text Embedding 3 SmallEmbeddings
openai/text-embedding-3-small - Text Embedding 3 LargeEmbeddings
openai/text-embedding-3-large - Text Embedding Ada 002Embeddings
openai/text-embedding-ada-002
Anthropic
- Claude Fable 5.1Messages
anthropic/claude-fable-5.1 - Claude Opus 5.5Messages
anthropic/claude-opus-5.5 - Claude Sonnet 5Messages
anthropic/claude-sonnet-5 - Claude Haiku 4.5Messages
anthropic/claude-haiku-4.5
More Anthropic models 6
- Claude Sonnet 4.5Messages
anthropic/claude-sonnet-4.5 - Claude Sonnet 4.6Messages
anthropic/claude-sonnet-4.6 - Claude Opus 4.5Messages
anthropic/claude-opus-4.5 - Claude Opus 4.6Messages
anthropic/claude-opus-4.6 - Claude Opus 4.7Messages
anthropic/claude-opus-4.7 - Claude Opus 4.8Messages
anthropic/claude-opus-4.8
DeepSeek
- DeepSeek V4.1 FlashChat Completions
deepseek/deepseek-v4.1-flash - DeepSeek V4 ProChat Completions
deepseek/deepseek-v4-pro - DeepSeek V4 Pro 0813Chat Completions
deepseek/deepseek-v4-pro-0813
Endpoints, tools and limitations
Use /v1/chat/completions, /v1/responses, /v1/messages or /v1/embeddings to match the endpoint listed for your model.
Chat Completions and Messages accept supported client-defined function tools. GPT-6 Sol and Luna require reasoning_effort: none for Chat Completions with tools. Responses supports text-only requests: tools, reasoning summaries and encrypted reasoning content stop before provider dispatch. A listed Codex model does not mean full Codex CLI compatibility.
Claude Fable 5.1 and Opus 5.5 use adaptive thinking. If you set thinking.type, use adaptive; forced tool choices (any or a specific tool) are not supported.
Only synchronous text generation and input-only embeddings at standard public pricing are qualified. Images, audio, built-in provider tools, batch, flex, fast and regional pricing modes are not qualified.
MiniRouter sends own-key requests directly to the provider using your encrypted key. Vercel AI Gateway is not in this request path. Your provider account must have access to the model. Unqualified requests stop before provider dispatch.
Set up a pool
- SaveName and encrypt a provider key.
- Group and orderGroup keys sharing provider funds, then set priority.
- EnableAcknowledge 3% before enabling a second key.
Provider keys are available to every account. Own-key mode remains off until you save your first enabled key or turn on a saved provider pool. Add and rotate credentials in Provider keys. A newly saved second key stays disabled until you enable it. Funding-account labels are customer-provided unless the provider confirms the scope.
Own-key mode applies per provider. Enable an OpenAI key and OpenAI requests use your keys; Anthropic, DeepSeek and other providers continue using MiniRouter credits unless you enable their own-key pools.
- Disable or revoke the last key
- That provider stays in own-key mode and its requests stop. Other providers are unchanged.
- Stop using one provider's keys
- That provider returns to MiniRouter keys at normal pricing. Other provider pools are unchanged.
- Stop using the last active pool
- All providers now use MiniRouter keys at normal pricing.
What switching can doAt most three upstream dispatches are authorized. An ambiguous timeout or a stream that started is not blindly replayed.
- Insufficient provider credit
- Try the next eligible independent key.
- Temporary rate limit
- Honor Retry-After, cool down the shared funding account, and try the next independent eligible key.
- Invalid or revoked key
- Stop using it until you replace or revalidate it.
- Model permission error
- Return without trying another key. The credential remains available for other models.
- Invalid request
- Return the error without switching keys.
- Ambiguous timeout or partial stream
- Do not replay work that the provider may already have billed.
Balances
- Provider reportedShown with its freshness when the provider offers a suitable balance API.
- UnavailableNo amount is invented. Confirmed credit failures can still move to another key.
- MiniRouter budgetSet an external-spend cap and low-balance threshold. They cover traffic MiniRouter sees, not direct requests.
When own keys cannot serve
If an enabled provider pool has no qualified eligible key, its requests stop without switching to MiniRouter credits. Providers whose pools are off or unconfigured use MiniRouter credits normally. Earlier own-key attempts may still appear on your provider invoice.
Receipts
Recent receipts in Provider keys keep provider reference usage, the applied 0% or 3% rate, collected fee and any waiver separate. The provider invoice remains separate.