Providers

Compare provider modalities, input media, batch transport, metadata fields, and webhook behavior across all seven adapters.

Batchwork exposes one lifecycle across seven provider-native batch APIs. The provider still controls models, pricing, limits, processing time, and supported request features.

Output modalities

Provider	Text	Embeddings	Image generation
OpenAI	Yes	Yes	Yes
Anthropic	Yes	No	No
Google Gemini	Yes	Yes	Yes
Groq	Yes	No	No
Mistral	Yes	Yes	No
Together AI	Yes	No	No
xAI	Yes	No	Yes

Pass options with --provider-options or --provider-options-file. Keys use the exact SDK-style camel-case spelling documented on each provider page. Batchwork merges only the selected provider branch; per-record non-null values override defaults with a shallow merge.

OpenAI, Anthropic, Google, Groq, Mistral, and xAI reject unknown keys before provider upload. Together accepts trusted passthrough options but rejects fields reserved by the canonical request envelope. Canonical settings and provider options that control the same wire field fail as conflicts instead of silently overriding one another.

Exact text option keys:

OpenAI common: forceReasoning, logprobs, promptCacheKey, promptCacheRetention, safetyIdentifier, serviceTier, systemMessageMode, user. Chat additionally accepts logitBias, maxCompletionTokens, metadata, parallelToolCalls, prediction, reasoningEffort, store, textVerbosity. Responses additionally accepts allowedTools, contextManagement, conversation, include, instructions, maxToolCalls, metadata, parallelToolCalls, previousResponseId, reasoningEffort, reasoningSummary, store, textVerbosity, truncation. Completions accepts echo, logitBias, logprobs, suffix, user.
Anthropic: cacheControl, container, contextManagement, disableParallelToolUse, effort, fallbacks, inferenceGeo, mcpServers, metadata, speed, taskBudget, thinking, toolStreaming.
Google: audioTimestamp, cachedContent, imageConfig, labels, mediaResolution, responseModalities, retrievalConfig, safetySettings, serviceTier, thinkingConfig.
Groq: parallelToolCalls, reasoningEffort, reasoningFormat, serviceTier, user.
Mistral: documentImageLimit, documentPageLimit, parallelToolCalls, reasoningEffort, safePrompt.
xAI: include, logprobs, previousResponseId, reasoningEffort, reasoningSummary, store, topLogprobs.
Together: provider-owned keys pass through except reserved custom_id, customId, input, messages, model, prompt, and stream.

Treat these options as trusted configuration, not end-user input. Provider APIs remain authoritative for model-specific and evolving constraints; Batchwork validates only constraints it can decide locally before upload.

This matrix describes generated output. Text requests can accept additional media inputs.

Text-request input media

Provider	Images	PDFs/files	Audio	Direct provider URLs
OpenAI	Yes	Responses PDFs/files	Supported forms	Images; PDFs for Responses
Anthropic	Yes	PDFs/documents	No	Images and PDFs
Google	Yes	Inline/provider files	Supported inline forms	Constrained Files and YouTube URLs
Groq	Yes	No	No	Images
Mistral	Yes	PDFs	No	HTTPS PDF URLs
Together AI	Yes	Inline PDF/text	Inline WAV/MP3	Generally downloaded and inlined
xAI	Yes	File URLs/provider IDs	No	Images and supported file URLs

Provider/model support applies. See Text and messages and Security.

Submission and results

Provider	Submission	Results
OpenAI	JSONL file upload	Output and error JSONL files
Anthropic	Inline Message Batch	Same-origin JSONL result URL
Google	Inline long-running operation	Inline operation response only
Groq	JSONL file upload	Output and error JSONL files
Mistral	JSONL file upload	Output and error JSONL files
Together AI	Upload initiation, presigned PUT, preprocessing	Output and error JSONL files
xAI	JSONL file upload	Paginated results endpoint

Batch-level metadata

The metadata= submission argument is not supported uniformly. Unsupported providers reject it locally before upload or batch creation.

Forwarded	Rejected
OpenAI, Groq, Mistral, Together AI	Anthropic, Google, xAI

Request-level provider metadata may still be available through provider_options. Google uses provider request metadata internally to preserve custom_id correlation.

Webhooks

OpenAI is the only provider with a native inbound batch-webhook helper. Every provider can use Batchwork-managed polling and signed outbound completion delivery through BatchPoller.

Credentials

See Configuration for API keys, endpoint overrides, and precedence. Each provider page documents its default API root and authentication style.