Supported providers
Responses provider usage preserves both input/cache details and output/reasoning details through normalization. Responses clients receive those details back, not just the three headline token counts.
Responses-format adapters preserve the client's full reasoning object and
explicit parallel_tool_calls setting. This includes Codex Responses-Lite's
context setting; provider-default routing does not remove client choices.
The OpenAI Codex adapter preserves non-message Responses input items, including additional_tools and tool history. This same-format preservation does not imply support for those items on other providers' API formats.
When supplied, OpenAI-compatible nested cache-write counts are normalized alongside cache reads. For the Codex Responses format, input_tokens_details.cache_write_tokens becomes the logged cache-write count; an explicit zero stays zero, while an absent count remains unknown. These figures feed cache evidence and cost accounting.
Each row is a distinct adapter. Where a provider offers both a pay-as-you-go API and a flat-rate coding subscription, they're listed as separate key types since they use different base URLs and pricing.
| Provider | Key type(s) | Auth | Notes |
|---|---|---|---|
| OpenAI | openai |
Bearer | GPT-4o/4.1, o1/o3, GPT-5.x |
| OpenAI Codex (ChatGPT) | openai-codex |
OAuth (device flow, in-app) | Uses your ChatGPT plan's included usage rather than per-token billing; cost tracked as $0 |
| Anthropic | anthropic |
x-api-key | Claude models |
| Anthropic (Claude Pro/Max) | anthropic_oauth |
Bearer setup-token | Paste a token from claude setup-token (run the Claude Code CLI's own command); flat-rate, no per-token cost |
| Google Gemini | google |
Query param | Gemini 1.5/2.x family |
| Groq | groq |
Bearer | Fast inference on open models |
| Mistral | mistral |
Bearer | Mistral/Codestral family |
| Cohere | cohere |
Bearer | Command family, Cohere's v2 chat API |
| DeepSeek | deepseek |
Bearer | deepseek-chat, deepseek-reasoner |
| Moonshot / Kimi | moonshot |
Bearer | Pay-as-you-go API |
| Moonshot / Kimi Code | moonshot_coding |
Bearer | Flat-rate coding subscription plan, $0 marginal cost |
| Wafer | wafer |
Bearer | Open models (GLM, Kimi, MiniMax, Qwen) on Wafer's serverless API; OpenAI-compatible |
| xAI (Grok) | xai |
Bearer | Grok family |
| z.ai (GLM) | zai_standard |
Bearer | Pay-as-you-go API |
| z.ai Coding Plan | zai_coding |
Bearer | Flat-rate coding subscription plan, $0 marginal cost |
Every adapter normalizes provider-specific quirks so the fallback chain "just works": each provider signals a context-window overflow differently (Anthropic returns a 400 with "prompt is too long" in the message; z.ai returns HTTP 200 with finish_reason: "model_context_window_exceeded"; OpenAI uses an explicit error code) and DodoRouter detects all of these and treats them the same way — as a fallback-eligible error, unless you've turned on "Skip fallback on context overflow". One known exception: Wafer reports an overflow with the same generic 400 it uses for any invalid request, so DodoRouter can't tell them apart — an oversized request on a Wafer step still falls back to the next step, but it isn't recognized as an overflow specifically, so the skip-fallback option doesn't apply to it. Cache-token reporting (for cost and cache-hit-rate tracking) is likewise normalized across each provider's own field names.