Provider Comparison

This library currently ships first-party adapters for Anthropic, OpenAI, and Google Gemini.

Capability Matrix

Provider	Selected seeded completion models	Streaming	Tool calling	Vision inputs	Session persistence	Notes
Anthropic	claude-sonnet-4-6, claude-haiku-4-5, claude-opus-4-6	Yes	Yes	Yes	Via Conversation + session stores	Anthropic cache read/write pricing is modeled separately, including block-level and request-level cache_control.
OpenAI	gpt-5.4, gpt-5.4-mini, gpt-5.4-nano, gpt-4o, gpt-4o-mini, o3	Yes	Yes	Yes	Via Conversation + session stores	Uses the stateless Responses API with store: false and library-owned history replay.
Google Gemini	gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite, gemini-3.1-pro-preview, gemini-3.1-flash-lite-preview	Yes	Yes	Yes	Via Conversation + session stores	Streaming uses the dedicated streamGenerateContent endpoint, and explicit caches are managed with client.googleCaches.

These are the checked-in seeded completion models, not the provider's full live catalog. Use client.models.listRemote({ provider }) when you want discovery against the provider's current model list.

Translation Differences

Concern	Anthropic	OpenAI	Gemini
System prompt handling	Lifted into system blocks	Flattened into top-level instructions	Lifted into systemInstruction
Assistant role name	assistant	assistant	model
Tool call payload	tool_use blocks	function_call.arguments JSON string with call_id	functionCall.args object
Tool result payload	tool_result block in a user turn	function_call_output item keyed by call_id	functionResponse part in a user turn
Streaming terminator	SSE close / message_stop	response.completed / stream close	SSE close on dedicated stream endpoint

Choosing a Provider

• Choose OpenAI when you want Responses-first model coverage, automatic prompt caching on supported models, and the broadest ecosystem compatibility.
• Choose Anthropic when long-context tool workflows or prompt caching behavior are central to the workload.
• Choose Gemini when you need a single provider surface that is comfortable with mixed text, vision, document, and audio inputs.

Operational Notes

• All three adapters normalize token usage into a shared UsageMetrics shape and estimate cost from the model registry.
• All three adapters map auth, rate-limit, context-window, and generic provider failures into typed LLMError subclasses.
• OpenAI and Gemini cached-read usage is priced separately from uncached input when the provider returns cached-token counts.
• Live-provider smoke tests can be executed with LIVE_TESTS=1 pnpm test:live after populating .env.