The Gateway normalizes every provider to OpenAI chat-completions semantics. This page covers what each provider does natively and which quirks the Gateway hides.

Tool-use matrix

Provider	Tool use	Parallel calls	Streaming deltas	`strict:true`	Forced `tool_choice`	Prompt caching	Known quirks
Anthropic Claude	✅	✅	✅ arg-by-arg	✅ server-side validation	✅ `"required"` + named	✅ `ephemeral`	`toolu_XXX` IDs are prefixed with `call_` by the Gateway for OpenAI-client compatibility.
Google Gemini	✅	✅	✅ whole-object (args don't chunk)	✅ server-side validation	✅ `"required"` + named	—	JSON schemas are sanitized before forwarding: `additionalProperties`, `$ref`, `$schema`, `$defs`, `strict`, and similar draft-meta keywords are dropped with a warning. IDs are synthesized as `call_<uuid>`. A `thinkingBudget` cap is applied to prevent reasoning tokens from eating the output budget.
OpenAI	✅	✅	✅ arg-by-arg	✅ passthrough	✅ `"required"` + named (passthrough)	✅ automatic (>1024 tok)	Passthrough.
xAI Grok	✅	✅	✅ arg-by-arg	✅ server-side validation	✅ `"required"` + named	—	Passthrough via OpenAI-compat endpoint.
DeepSeek (chat)	✅	✅	✅ arg-by-arg	✅ server-side validation	✅ `"required"` + named	—	`deepseek-reasoner` / `deepseek-r1` do not support tools — returns `tool_unsupported_for_model`. Use `deepseek-chat`.
Mistral	✅	✅	✅ arg-by-arg	✅ server-side validation	✅ `"required"` + named	—	Passthrough.
Kimi (Moonshot)	✅	provider-dependent	✅	✅ server-side validation	✅ `"required"` + named	—	Via OpenRouter.
Llama (Meta)	✅	provider-dependent	✅	✅ server-side validation	provider-dependent	—	Via OpenRouter. Tool support varies per Llama variant.
MiniMax (M2.7)	✅	✅	✅	✅ server-side validation	✅ `"required"` + named	—	MiniMax's hosted API is OpenAI-compatible including `tools` / `tool_choice` / `tool_calls`. The model serves XML-format tool calls internally but the wire format is standard OpenAI.

Reasoning models and tool use

Model	Tools allowed?
`gpt-5.2` (OpenAI o-series derivative)	✅
`claude-opus-4-*`	✅
`grok-4-1-fast-reasoning`	✅
`deepseek-reasoner` / `deepseek-r1`	❌ — rejected with `tool_unsupported_for_model`.

JSON Schema subset

The Gateway validates parameters as a JSON Schema Draft 2020-12 subset with type:"object" at the root. Supported keywords:

type, properties, required, description, enum
items, minItems, maxItems
oneOf, anyOf, allOf
pattern, minLength, maxLength
minimum, maximum
format (validation is lenient: uuid, date, date-time checked loosely)
additionalProperties (Claude/OpenAI/Grok/DeepSeek/Mistral; dropped for Gemini)

When a keyword isn't supported by the target provider, the Gateway either drops it with a warning (Gemini) or validates server-side.

Fallback behavior with tool use

When you pass fallback: ["gpt", "gemini"], the Gateway will try fallback providers only before any tool-call delta has been emitted. Once a tool call has streamed, no mid-stream failover is attempted — tool-call IDs and schemas wouldn't align across providers.

{
  "model": "claude",
  "fallback": ["gpt", "gemini"],
  "tools": [ /* ... */ ]
}

Anthropic 529 before any delta → try gpt.
Anthropic connection drop after first tool-call delta → surface tool_provider_error + [DONE].

Guardrails with tool use

Guardrails run on assistant text content only. Tool-call arguments are skipped — they're structured JSON often containing emails, UUIDs, IDs, and search terms that would false-positive against PII rules.

Token usage and cached tokens

The Gateway normalizes usage reporting:

prompt_tokens counts input tokens including cached-hit tokens, following OpenAI convention. It does not include cache-creation tokens (the first-write turn on Anthropic ephemeral).
completion_tokens counts the assistant output, including tool_calls arguments.
prompt_tokens_details.cached_tokens is populated for Anthropic (cache_read_input_tokens) and OpenAI (native cached_tokens). Other providers report 0.

`tool_choice` shapes

Every provider above accepts the two OpenAI-style forced shapes:

// Force *some* tool to be called — model picks which one
"tool_choice": "required"
 
// Force a *specific* tool by name — must match one of tools[].function.name
"tool_choice": { "type": "function", "function": { "name": "search_code" } }

A bad name in the named shape returns tool_choice_invalid.

Provider Compatibility