Channel capability matrix
Check credential types, client protocols, operations, route modes, model discovery, and managed proxy support by channel.
Capability scope §
This page describes route capabilities declared by GPT-Load. It does not mean that every model or upstream supports each capability.
N means Native execution in the target protocol; C means Converted by the gateway; N/C depends on the model. N* supports only a subset of native resource operations, detailed below the table. — means no route is currently declared.
Credentials and core protocols §
| Channel | Credential | Chat Completions | Responses Create | Anthropic Messages | Gemini Generate |
|---|---|---|---|---|---|
| Jev | API Key | — | — | — | — |
| Cerebras | API Key | N | C | C | C |
| Mistral | API Key | N | C | C | C |
| Nebius | API Key | N | C | C | C |
| Parasail | API Key | N | C | C | C |
| Wafer | API Key | N | C | C | C |
| Hugging Face | API Key | N | C | C | C |
| Cohere | API Key | — | — | — | — |
| OpenAI | API Key | N | N | C | C |
| Codex | OAuth / OAuth JSON | C | N | C | C |
| Claude | OAuth / OAuth JSON | C | C | N | C |
| Antigravity | OAuth / OAuth JSON | C | C | C | N |
| Grok | Device authorization / OAuth JSON | C | N | C | C |
| Anthropic | API Key | C | C | N | C |
| Google Gemini | API Key | C | C | C | N |
| Azure OpenAI | API key / Entra | C | C | C | C |
| AWS Bedrock | Bedrock API key / AWS credentials | C | C | C | C |
| Google Vertex AI | Service account JSON | C | C | C | N/C |
| DeepSeek | API Key | N | N | N | C |
| Moonshot AI | API Key | N | C | C | C |
| SiliconFlow | API Key | N | C | C | C |
| Zhipu AI | API Key | N | C | C | C |
| Alibaba Cloud Bailian | API Key | N | C | C | C |
| Volcengine Ark | API Key | N | C | C | C |
| OpenRouter | API Key | N | N | C | C |
| Cline | API Key | N | C | C | C |
| Groq | API Key | N | C | C | C |
| xAI | API Key | N | N | C | C |
| OpenCode Go | API Key | N | N | N | — |
| OpenCode Zen | API Key | N | N | N | — |
| GPT-Load | AccessKey + base URL | N | N | N | N |
| New API | API key + base URL | N | N | N | N |
| CLIProxyAPI | API key + base URL | N | N | N | N |
| Sub2API | API key + base URL | N | N | N | N |
| OpenAI Compatible | API key + custom endpoint | N | C | C | C |
Route modes are declared per protocol and operation. One channel can contain both Native and Converted routes.
Additional operations §
| Channel | Image generation/editing | OpenAI Embeddings | Anthropic CountTokens | Gemini CountTokens | Responses Input Tokens | Responses resource APIs | Model discovery | Managed proxy |
|---|---|---|---|---|---|---|---|---|
| Jev | — | — | — | — | — | — | ✓ | ✓ |
| Cerebras | — | — | — | — | — | — | ✓ | ✓ |
| Mistral | — | N | — | — | — | — | ✓ | ✓ |
| Nebius | — | N | — | — | — | — | ✓ | ✓ |
| Parasail | — | — | — | — | — | — | ✓ | ✓ |
| Wafer | — | — | — | — | — | — | ✓ | ✓ |
| Hugging Face | — | — | — | — | — | — | ✓ | ✓ |
| Cohere | — | — | — | — | — | — | — | ✓ |
| OpenAI | N | N | C | C | N | N | ✓ | ✓ |
| Codex | N | — | C | C | N | — | ✓ | ✓ |
| Claude | — | — | N | C | C | — | ✓ | ✓ |
| Antigravity | C | — | C | N | C | — | ✓ | ✓ |
| Grok | — | — | C | C | N | — | ✓ | ✓ |
| Anthropic | — | — | N | C | C | — | ✓ | ✓ |
| Google Gemini | C | C | C | N | C | — | ✓ | ✓ |
| Azure OpenAI | — | — | — | — | — | — | ✓ | — |
| AWS Bedrock | — | — | — | — | — | — | ✓ | — |
| Google Vertex AI | — | — | — | — | — | — | ✓ | — |
| DeepSeek | — | — | — | — | — | — | ✓ | ✓ |
| Moonshot AI | — | — | — | — | — | — | ✓ | ✓ |
| SiliconFlow | — | N | — | — | — | — | ✓ | ✓ |
| Zhipu AI | — | — | — | — | — | — | ✓ | ✓ |
| Alibaba Cloud Bailian | — | — | — | — | — | — | ✓ | ✓ |
| Volcengine Ark | — | — | — | — | — | — | ✓ | ✓ |
| OpenRouter | — | N | — | — | — | — | ✓ | ✓ |
| Cline | — | — | — | — | — | — | ✓ | ✓ |
| Groq | — | — | — | — | — | — | ✓ | ✓ |
| xAI | — | — | — | — | — | — | ✓ | ✓ |
| OpenCode Go | — | — | — | — | — | — | ✓ | ✓ |
| OpenCode Zen | — | — | — | — | — | — | ✓ | ✓ |
| GPT-Load | N | N | N | N | N | N | ✓ | ✓ |
| New API | N | N | — | — | — | N* | ✓ | ✓ |
| CLIProxyAPI | N | — | N | N | — | N* | ✓ | ✓ |
| Sub2API | N | — | N | N | — | N* | ✓ | ✓ |
| OpenAI Compatible | N | N | — | — | — | — | ✓ | ✓ |
For Images, N means native generation and editing. C means Gemini and Antigravity support only single-image, non-streaming, Base64 generation conversion, not editing. Responses resource interfaces cover retrieval, deletion, cancellation, input items, compaction, and namespace passthrough.
For N* in the Responses resource column, CLIProxyAPI and New API support Compact and Responses namespace extension passthrough; Sub2API supports Compact only. All three lack response retrieval, deletion, cancellation, input items, and Responses Input Tokens; the GPT-Load channel supports those operations. Extension passthrough does not guarantee upstream implementation, and upstream error statuses are preserved.
CLIProxyAPI, New API, and GPT-Load natively support POST /v1/alpha/search; the access key must allow openai-responses. Model mappings and search payloads are forwarded, preserving response content and status. Search does not count as model token usage. Verify that the upstream account and service allow search.
Cohere, SiliconFlow, OpenAI Compatible, New API, and GPT-Load support native text-only Rerank at POST /v1/rerank. Streaming and protocol conversion are not supported; the AccessKey must allow the rerank protocol.
Google Gemini, New API, and GPT-Load support native Gemini embeddings (embedContent/batchEmbedContents) using gemini-embeddings. Google Gemini's OpenAI Embeddings entry in the table is a text conversion route and does not accept token-ID input.
Jev and OpenRouter natively support decisions. The Codex subscription channel supports the independent codex-live voice protocol. Jev and Cohere have no conversation routes and should not be used as ordinary chat channels.
The Mistral channel natively supports OCR, FIM completions, audio transcription, speech synthesis, text and chat moderation, classification, voice management, and WebSocket realtime transcription. The AccessKey must allow the independent mistral protocol. Chat Completions and Embeddings still use their OpenAI-compatible protocols.
Model discovery means the management plane uses a credential to discover upstream models. Data-plane model lists return configured models visible to the current AccessKey and do not query upstream in real time.
Managed proxy means a proxy injected by GPT-Load at credential, Group, or global scope. Provider SDK behavior that reads environment variables is outside this contract.
Validation levels §
| Level | Meaning | Current public evidence |
|---|---|---|
| Automated contract | Channel declarations, RouteMode values, adapter capabilities, and automated tests agree | Covers the routing contracts of all 35 channels listed here |
| Artifact smoke | A released binary or image traverses the full product path, while the upstream may be a controlled substitute | The release image covers OpenAI Compatible Chat through to a fake upstream |
| Live upstream verified | A specific version, channel, operation, and model are verified with a real credential | No centralized, individually traceable live E2E records are currently published for all 35 channels |
Upstreams change APIs, models, and account policies. A verification record without a version, date, operation, model, and evidence link is not marked Live upstream verified.
Known limitations §
- The actual capabilities of OpenAI Compatible depend on the relay you configure. The matrix only states which requests GPT-Load can construct and accept.
- Google Vertex AI chooses Native or Converted Gemini routing by model.
- OpenCode Go and OpenCode Zen support only the native protocols available for the upstream model. Responses creation is stateless; Gemini and Responses resource endpoints are not supported.
- Mistral native endpoints do not include files, batch, agents, conversations, or browser sessions using temporary tokens.
- A protocol route does not guarantee that a model supports tools, reasoning, vision, or other model-level features.
- Azure OpenAI, AWS Bedrock, and Google Vertex AI do not support GPT-Load managed proxies.