Client parameter compatibility

This document is a public reference for the parameter compatibility of the OpenAI, Anthropic, and Gemini-like client SDKs in gpt2giga. It reflects the current structure of the source code:

• OpenAI routers: gpt2giga/routers/openai/
• Anthropic routers: gpt2giga/routers/anthropic/
• Gemini routers: gpt2giga/routers/gemini/
• shared request policies: gpt2giga/common/client_params.py
• OpenAI request classification: gpt2giga/protocol/request/params.py
• Anthropic request classification: gpt2giga/protocol/anthropic/params.py
• Gemini adapter: gpt2giga/protocols/gemini/

Other client families are not covered by this compatibility check.

Compatibility statuses

Status	Meaning
supported	The parameter affects the request or response and is covered by tests.
accepted_ignored	The parameter is accepted for SDK compatibility but is not sent upstream.
rejected	The request has an unexecutable shape, for example a missing required input or an extra_body that is not an object. Optional client feature flags do not use this status.
not_applicable	The option relates to client-side transport configuration, not to a server-side request body parameter.

SDK transport options

base_url, api_key, timeout, retry settings, a custom http_client, proxy configuration, and low-level transport settings are client-side SDK options. gpt2giga does not assign them server-side semantics.

Credentials and transport headers are not forwarded to GigaChat as arbitrary upstream metadata. This applies to Authorization, x-api-key, cookies, host, content and transfer headers, x-stainless-*, openai-*, and anthropic-*. To intentionally pass GigaChat authorization, use the separate GPT2GIGA_PASS_TOKEN=True mode.

extra_headers and extra_query

extra_headers from an SDK arrives at the server as regular HTTP headers. gpt2giga moves safe headers into the request-scoped contextvars of the GigaChat SDK:

• x-request-id
• x-session-id
• x-service-id
• x-operation-id
• x-client-id
• x-trace-id
• x-agent-id

Other safe custom headers are passed through custom_headers_cvar. The diagnostic x-correlation-id and traceparent can also be passed as custom headers. Authorization, x-api-key, transport headers, and the SDK-internal x-stainless-*, openai-*, anthropic-* headers remain blocked.

The SDK URL contextvars, for example chat_url_cvar and chat_completions_url_cvar, are not populated from extra_headers.

extra_query does not forward arbitrary query parameters upstream by default: the list of allowed upstream query parameters is empty.

extra_body

The OpenAI and Anthropic SDKs usually merge extra_body with the outgoing JSON body as top-level fields. HTTP clients that work directly can also send a literal extra_body object. gpt2giga handles both forms.

For OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages, the extra_body object is moved into GigaChat additional_fields in full. SDK-style unknown top-level fields that appear after the client expands extra_body are handled the same way.

Known unsupported optional client parameters are accepted and ignored if sent as top-level fields: for example logprobs, audio, container, or mcp_servers. previous_response_id is supported for OpenAI Responses in GigaChat v2 mode and is mapped to storage.thread_id; in Responses v1 mode it is accepted and ignored. If the same key is explicitly placed inside a literal extra_body, gpt2giga passes it to additional_fields, and the GigaChat upstream determines the final support.

OpenAI Embeddings accepts and ignores extra_body, unknown top-level fields, and dimensions; at the moment there are no GigaChat-specific allowed fields for embeddings.

OpenAI body parameters

Endpoint	Supported
Chat Completions	model, messages, stream, temperature, top_p, max_tokens, max_completion_tokens, stop, function tools, functions, function_call, supported tool_choice, built-in tools in GigaChat v2 mode (web_search*, code_interpreter, image_generation / image_generate, url_content_extraction, model_3d_generate), response_format, reasoning, reasoning_effort, extra_body passthrough
Responses	model, input, instructions, stream, temperature, top_p, max_output_tokens, function tools, built-in tools in GigaChat v2 mode (web_search*, code_interpreter, image_generation / image_generate, url_content_extraction, model_3d_generate; normalized output items and stream progress events are currently built for web_search* and image_generation / image_generate), supported tool_choice, text.format, response_format, reasoning, reasoning_effort, extra_body passthrough
Embeddings	input, model, dimensions, encoding_format, user, extra_headers, extra_query
Models	GET /models, GET /models/{model}

Structured output is supported through json_schema. Schema-less JSON mode (response_format.type=json_object in OpenAI or Gemini responseMimeType=application/json without responseJsonSchema / responseSchema) is rejected, because the GigaChat upstream does not support a separate JSON mode.

With GPT2GIGA_DISABLE_REASONING=True, the proxy accepts reasoning and reasoning_effort but does not pass them to the upstream payload sent to GigaChat.

With GPT2GIGA_DISABLE_BUILTIN_TOOL_MAPPING=True, the proxy accepts provider built-in tools for compatibility but does not map or send them to GigaChat as executable tools. User function tools continue to work.

OpenAI metadata fields such as user, metadata, service_tier, safety_identifier, seed, prompt_cache_key, and prompt_cache_retention are accepted and ignored where they are classified.

Unsupported optional OpenAI parameters are accepted and ignored. Examples: logprobs, top_logprobs, logit_bias, audio output, prediction, web_search_options, built-in tools outside GigaChat v2 mode, n > 1, parallel_tool_calls=true, stored completions requests, conversation, and previous_response_id in Responses v1 mode. /chat/completions v1 remains a supported compatibility route, but new tool/built-in-tool capabilities evolve for GigaChat v2/chat/completions.

Anthropic body parameters

Endpoint	Supported
Messages	model, messages, system, max_tokens, stream, temperature, top_p, stop_sequences, local function tools, Anthropic provider tools in GigaChat v2 mode (web_search*, web_fetch* as url_content_extraction, code_execution* as code_interpreter), tool_choice values auto/none/forced tool, thinking, output_config.format, output_format, extra_body passthrough
Count Tokens	model, messages, system, tools, structured-output schema text, compatible message content validation
Models	GET /models, GET /models/{model_id}, when the request contains Anthropic SDK headers, for example anthropic-version

Anthropic metadata, service_tier, top_k, beta headers, and betas are accepted and ignored where they are classified.

Unsupported optional Anthropic parameters or features are accepted and ignored. Examples: container, context_management, mcp_servers, unsupported provider tools (advisor, tool_search, mcp_toolset, memory, bash, text_editor, computer), computer use, document/file content blocks, container uploads, search result blocks, citations, and thinking/redacted_thinking input blocks.

The Anthropic Message Batches code exists, but the public router is not mounted until batch operation support appears in the GigaChat SDK or backend.

Gemini body parameters

Gemini-like operation routes are available at the root, under /v1, /v2, and /v1beta, for example /v1/models/{model}:generateContent. The Gemini SDK/CLI paths /v1/v1beta/... and /v2/v1beta/... are also supported when the client itself adds /v1beta to a versioned base URL. /v1 and /v1/v1beta force the GigaChat v1 backend contract, /v2 and /v2/v1beta force the GigaChat v2 backend contract. Root paths without /v1 or /v2, including /v1beta/..., select the backend by GPT2GIGA_GIGACHAT_API_MODE=v1|v2.

Endpoint	Supported
Generate Content	contents, systemInstruction, generationConfig.temperature, generationConfig.topP, generationConfig.maxOutputTokens, generationConfig.stopSequences, generationConfig.seed, generationConfig.presencePenalty, generationConfig.frequencyPenalty, function tools.functionDeclarations, Gemini provider tools in GigaChat v2 mode (googleSearch / googleSearchRetrieval as web_search, urlContext as url_content_extraction, codeExecution as code_interpreter), basic toolConfig.functionCallingConfig, text/image/file parts
Stream Generate Content	The same fields as Generate Content; the response is returned as Gemini GenerateContentResponse SSE chunks.
Count Tokens	Text parts of contents, systemInstruction, and function declaration names/descriptions.
Embeddings	content.parts[].text, requests[].content.parts[].text for batch embeddings, outputDimensionality accepted as compatibility metadata.
Models	GET /v1beta/models, GET /v1/v1beta/models, GET /v2/v1beta/models and {model} variants; the shared /models, /v1/models, /v2/models return the Gemini form for Google/Gemini requests, for example with X-Goog-Api-Client

Gemini safetySettings, cachedContent, serviceTier, store, and unsupported generationConfig subfields are accepted and kept in normalized extensions for diagnostics, but are not passed to GigaChat as executable parameters and are not applied by the proxy. Non-function provider tools that do not match the GigaChat SDK built-in tools (fileSearch, googleMaps, computerUse, MCP, RAG/retrieval/Vertex tools) are also kept for diagnostics only. The full list of canonical built-in tools and provider aliases is described in Built-in tools.

The Gemini Files and Batches code exists, but the public router is not mounted until file/batch execution is verified end-to-end.

Route scope

The mounted OpenAI, Anthropic, LiteLLM, and Gemini operation routes are available at the root, under /v1, and under /v2 via the application router. Gemini-compatible routes are also available under /v1beta, /v1/v1beta, and /v2/v1beta. /v1 always selects the GigaChat v1 backend contract, /v2 always selects the GigaChat v2 backend contract, and the root without a versioned prefix follows GPT2GIGA_GIGACHAT_API_MODE=v1|v2. The prepared but disabled OpenAI Files/Batches, Anthropic Message Batches, and Gemini Files/Batches routes are intentionally excluded from the OpenAPI schema by default.