API pricing

What breaks when you switch AI APIs

What breaks when you move an app from one AI API to another, at 10 providers: strict JSON schemas, parallel tool calls, streamed tool arguments, reasoning with tools, and caching of tool definitions. Each answer is yes, partly, no or not stated, with the provider’s own words and the date we read them, or with our own test call and its date.

Before you switch

Strict JSON schemas: partly at DeepSeek, Alibaba Cloud (Qwen), Z.ai (GLM). Code that parses answers without checks starts failing where a provider promises only valid JSON, not your schema.
Parallel tool calls: partly at Groq; not stated at DeepSeek, Z.ai (GLM). An agent built around several calls per turn gets one call per turn: more round trips, more input tokens, slower runs.
Streamed tool arguments: no at xAI, Groq, Mistral; partly at Z.ai (GLM). An interface that shows arguments as they are written sees nothing until the whole call is ready.
Reasoning with tools: partly at Cerebras; not stated at Mistral. Switching reasoning on can rule out some tool settings or output formats, or the other way round.
Cache covers tool definitions: not stated at Google Gemini, DeepSeek, Alibaba Cloud (Qwen), xAI, Mistral, Z.ai (GLM). Where the cache does not cover tool definitions, a long tool list is paid as fresh input on every request.

At a glance

Compare API features visually

Provider	Strict JSON schemas	Parallel tool calls	Streamed tool arguments	Reasoning with tools	Cache covers tool definitions
Anthropic	Yes PublishedObserved	Yes Published	Yes Published	Yes Published	Yes Published
OpenAI	Yes PublishedObserved	Yes Published	Yes Published	Yes Published	Yes Published
Google Gemini	Yes PublishedObserved	Yes Published	Yes Published	Yes Published	Not stated Published
DeepSeek	Partly PublishedObserved	Not stated Unknown	Yes Observed	Yes Published	Not stated Published
Alibaba Cloud (Qwen)	Partly PublishedObserved	Yes Published	Yes Observed	Yes Published	Not stated Published
xAI	Yes Published	Yes Published	No Published	Yes Published	Not stated Unknown
Groq	Yes Published	Partly Published	No Observed	Yes Published	Yes Published
Cerebras	Yes Published	Yes Published	Yes Observed	Partly Published	Yes Published
Mistral	Yes Published	Yes Published	No Observed	Not stated Published	Not stated Published
Z.ai (GLM)	Partly PublishedObserved	Not stated Published	Partly Published	Yes Published	Not stated Published

Published Published: quoted from the provider’s documentation, with a link and the date we read it. Observed Observed: our own test call, with its date. Unknown Unknown: not stated in the documentation we read.

Strict JSON schemas

The answer, or a tool call’s arguments, is guaranteed to match the JSON Schema you send. What breaks: Code that parses answers without checks starts failing where a provider promises only valid JSON, not your schema.

Provider	Support	In the provider’s words, or in our test call
Anthropic	Yes	Published Yes — The schema goes in `output_config.format` with type “`json_schema`”. “Structured outputs guarantee schema-compliant responses through constrained decoding:” provider documentation, read 17 Oct 2026 Published Yes — “The request carries the schema in `output_config.format` with type: "`json_schema`". SDK helpers set this for you.” provider documentation, read 17 Oct 2026 Published Yes · Claude Haiku 5.5 — From the list of supported models on the page. “claude-opus-4-5-20251101 - claude-haiku-5-5 - claude-haiku-4-5-20251001” provider documentation, read 17 Oct 2026 Observed Yes · claude-haiku-5-5 — Our test call on 17 Oct 2026: the API accepted a JSON Schema output format and the answer matched the schema. One call shows what the API did that day, not a guarantee. Published Yes — Strict schema for tool arguments. “Setting strict: true on a tool definition guarantees Claude's tool inputs match your JSON Schema by constraining the model's token sampling to schema-valid outputs (a technique called grammar-constrained sampling).” provider documentation, read 17 Oct 2026
OpenAI	Yes	Published Yes — “Structured Outputs is a feature that ensures the model will always generate responses that adhere to your supplied JSON Schema” provider documentation, read 17 Oct 2026 Observed Yes · gpt-5.4-mini — Our test call on 17 Oct 2026: the API accepted a strict JSON Schema response format and the answer matched the schema. One call shows what the API did that day, not a guarantee.
Google Gemini	Yes	Published Yes — “You can configure Gemini models to generate responses that adhere to a provided JSON Schema.” provider documentation, read 17 Oct 2026 Published Yes — Strict schema together with tools. “Gemini 3 lets you combine Structured Outputs with built-in tools, including Grounding with Google Search , URL Context , Code Execution , File Search , and Function Calling .” provider documentation, read 17 Oct 2026 Observed Yes · gemini-3.8-flash — Our test call on 17 Oct 2026: the API accepted a JSON Schema response format and the answer matched the schema. One call shows what the API did that day, not a guarantee.
DeepSeek	Partly	Published Partly — JSON mode (`json_object`); the page does not describe a JSON Schema response format. “DeepSeek provides JSON Output to ensure the model outputs valid JSON strings.” provider documentation, read 17 Oct 2026 Observed No · deepseek-flash — Our test call on 17 Oct 2026: the API rejected a JSON Schema response format with HTTP 400 (“This `response_format` type is unavailable now”); JSON mode is what remains. One call shows what the API did that day, not a guarantee. Published Yes — Strict schema for tool arguments (strict mode, marked Beta). “In strict mode, the model strictly adheres to the format requirements of the Function's JSON schema when outputting a tool call, ensuring that the model's output complies with the user's definition. It is supported by both thinking and non-thinking mode.” provider documentation, read 17 Oct 2026
Alibaba Cloud (Qwen)	Partly	Published Partly — “Usage Structured output supports two modes: JSON Object and JSON Schema.” provider documentation, read 17 Oct 2026 Published Partly — Columns of the table: JSON Object mode, JSON Schema mode. “Strictly follows schema No Yes Supported models Most Qwen models Only selected qwen-plus models” provider documentation, read 17 Oct 2026 Observed Yes · qwen3.8-max — Our test call on 17 Oct 2026: qwen3.8-max accepted a strict JSON Schema response format and answered to the schema, although the provider’s own table lists JSON Schema mode only for selected qwen-plus models. One call shows what the API did that day, not a guarantee.
xAI	Yes	Published Yes — “By setting `response_format.type` to "`json_schema`" and providing your schema under `response_format.json_schema` , you can define exactly what structured output the model should return.” provider documentation, read 17 Oct 2026 Published Yes — Tool arguments are always strict; the strict flag is implied. “When you define tools, [xAI] models will always generate tool call arguments that strictly conform to the tool’s input JSON Schema (the strict flag is implicitly always true ).” provider documentation, read 17 Oct 2026; provider name corrected in brackets
Groq	Yes	Published Yes — “Guarantee model responses strictly conform to your JSON schema for reliable, type-safe data structures.” provider documentation, read 17 Oct 2026 Published Yes — “With strict: true, the model uses constrained decoding to guarantee that the output will always match your schema exactly.” provider documentation, read 17 Oct 2026
Cerebras	Yes	Published Yes — “strict: true uses constrained decoding to guarantee that output conforms to schemas in the supported JSON Schema subset.” provider documentation, read 17 Oct 2026 Published Yes · gpt-oss-120b — Columns: text, JSON object, JSON Schema, strict, availability. “gpt-oss-120b · Yes · Yes · Yes · Yes · Shared Inference” provider documentation, read 17 Oct 2026
Mistral	Yes	Published Yes — “We offer a reliable method to obtain structured output in your desired format. Our system includes a built-in mode for JSON output, along with the capability to use custom structured outputs.” provider documentation, read 17 Oct 2026
Z.ai (GLM)	Partly	Published Partly — The page describes only JSON mode (`json_object`). “Structured output (JSON mode) ensures that AI returns JSON data conforming to predefined formats, providing reliable guarantees for programmatic processing of AI outputs.” provider documentation, read 17 Oct 2026 Observed Yes · glm-5.3 — Our test call on 17 Oct 2026: glm-5.3 accepted a strict JSON Schema response format and answered to the schema; the documentation describes only JSON mode. One call shows what the API did that day, not a guarantee.

Parallel tool calls

The model can ask for several tool calls in one turn. What breaks: An agent built around several calls per turn gets one call per turn: more round trips, more input tokens, slower runs.

Provider	Support	In the provider’s words, or in our test call
Anthropic	Yes	Published Yes — “By default, Claude may call multiple tools in a single response.” provider documentation, read 17 Oct 2026
OpenAI	Yes	Published Yes — “On supported models beginning with GPT-5, functions can be called in parallel when built-in tools are also available.” provider documentation, read 17 Oct 2026 Published Yes — Parallel calls are on by default; this switches them off. “You can prevent this by setting `parallel_tool_calls` to false, which ensures exactly zero or one tool is called.” provider documentation, read 17 Oct 2026
Google Gemini	Yes	Published Yes — “The model supports calling multiple functions in a single turn ( parallel function calling ) and in sequence ( compositional function calling ).” provider documentation, read 17 Oct 2026
DeepSeek	Not stated	Unknown Not stated — The Tool Calls and Chat Completion pages do not mention parallel tool calls, and a test call cannot force the model to make two calls.
Alibaba Cloud (Qwen)	Yes	Published Yes — “you can set the `parallel_tool_calls` request parameter to true when you make a Function Calling . Note Parallel tool calling is suitable for tasks that have no dependencies.” provider documentation, read 17 Oct 2026
xAI	Yes	Published Yes — Parallel calls are on by default; this switches them off. “Disable with `parallel_tool_calls`: false in your request.” provider documentation, read 17 Oct 2026
Groq	Partly	Published No · gpt-oss-120b — Columns: tool use, parallel tool use, JSON mode, built-in tools. No parallel calls for this model. “openai/gpt-oss-120b · Yes · No · Yes · Yes” provider documentation, read 17 Oct 2026 Published Yes · qwen3.8-27b — Columns: tool use, parallel tool use, JSON mode, built-in tools. “qwen/qwen3.8-27b · Yes · Yes · Yes · No” provider documentation, read 17 Oct 2026
Cerebras	Yes	Published Yes — “Strict mode works with parallel tool calling.” provider documentation, read 17 Oct 2026 Published Yes · gpt-oss-120b — Columns: `tool_choice`, strict mode, parallel calls, multi-turn calling, availability. “gpt-oss-120b · none, auto, required, named function · Yes · Yes · Yes · Shared Inference” provider documentation, read 17 Oct 2026
Mistral	Yes	Published Yes — “And `parallel_tool_calls` to specify whether parallel tool calling is allowed. - true: default mode. The model decides if it uses parallel tool calls or not.” provider documentation, read 17 Oct 2026
Z.ai (GLM)	Not stated	Published Not stated — The page does not mention parallel calls. “`tool_choice`: Controls function calling strategy, default is auto (only supports auto)” provider documentation, read 17 Oct 2026

Streamed tool arguments

Tool-call arguments arrive in pieces while the model writes them. What breaks: An interface that shows arguments as they are written sees nothing until the whole call is ready.

Provider	Support	In the provider’s words, or in our test call
Anthropic	Yes	Published Yes — Arguments arrive as `input_json_delta` events, one whole key and value at a time. “Note: Current models only support emitting one complete key and value property from input at a time.” provider documentation, read 17 Oct 2026
OpenAI	Yes	Published Yes — “Streaming can be used to surface progress by showing which function is called as the model fills its arguments, and even displaying the arguments in real time.” provider documentation, read 17 Oct 2026
Google Gemini	Yes	Published Yes — “Tool arguments can be streamed as partial arguments using arguments . You must aggregate these deltas to reconstruct the complete tool calls” provider documentation, read 17 Oct 2026
DeepSeek	Yes	Observed Yes · deepseek-flash — Our test call on 17 Oct 2026: the tool-call arguments arrived in 33 pieces in the stream. One call shows what the API did that day, not a guarantee.
Alibaba Cloud (Qwen)	Yes	Observed Yes · qwen3.8-max — Our test call on 17 Oct 2026: the tool-call arguments arrived in 14 pieces in the stream. One call shows what the API did that day, not a guarantee.
xAI	No	Published No — “With streaming, the function call is returned in whole in a single chunk, not streamed across chunks.” provider documentation, read 17 Oct 2026
Groq	No	Observed No · gpt-oss-120b — Our test call on 17 Oct 2026: the arguments (about 60 characters) arrived in one piece. One call shows what the API did that day, not a guarantee.
Cerebras	Yes	Observed Yes · gpt-oss-120b — Our test call on 17 Oct 2026: the tool-call arguments arrived in 3 pieces in the stream. One call shows what the API did that day, not a guarantee.
Mistral	No	Observed No · mistral-large-4-0 — Our test call on 17 Oct 2026: the arguments (about 60 characters) arrived in one piece. One call shows what the API did that day, not a guarantee.
Z.ai (GLM)	Partly	Published Partly — Switched on with the `tool_stream` parameter. “Streaming output for tool call responses allows developers to stream tool usage parameters without buffering or JSON validation when calling chat.completions, reducing call latency and providing better user experience.” provider documentation, read 17 Oct 2026 Published Partly · glm-5.3 — The page limits `tool_stream` to glm-5 and does not mention glm-5.3. “limited to glm-5” provider documentation, read 17 Oct 2026

Reasoning with tools

Thinking or reasoning works in the same request as tool calls. What breaks: Switching reasoning on can rule out some tool settings or output formats, or the other way round.

Provider	Support	In the provider’s words, or in our test call
Anthropic	Yes	Published Yes — “Interleaved thinking lets Claude think between tool calls within a single assistant turn, reasoning about each tool result before determining what to do next.” provider documentation, read 17 Oct 2026
OpenAI	Yes	Published Yes — “Our reasoning models like gpt-5.5 and gpt-5.4 support interleaved thinking, where the model is able to generate visible output tokens before and in between thinking, and is able to think in between tool calls.” provider documentation, read 17 Oct 2026
Google Gemini	Yes	Published Yes — “Function calling with thinking models Gemini 3 series models use an internal "thinking" process that improves function calling.” provider documentation, read 17 Oct 2026
DeepSeek	Yes	Published Yes — “From DeepSeek-V3.2, the API supports tool use in the thinking mode.” provider documentation, read 17 Oct 2026
Alibaba Cloud (Qwen)	Yes	Published Yes — Tools work in thinking mode; `tool_choice` “required” does not. “cannot guarantee that a tool is always called, and in thinking mode "required" is not supported” provider documentation, read 17 Oct 2026
xAI	Yes	Published Yes — Reasoning cannot be switched off on grok-4.7; low effort is the recommended setting for tool calls. “"low" Uses some reasoning tokens, but still fast Latency-sensitive agentic use and simple tool calling.” provider documentation, read 17 Oct 2026
Groq	Yes	Published Yes — Reasoning and tools work together; the raw reasoning format is not allowed with tools. “Must be set to either parsed or hidden when using tool calling or JSON mode.” provider documentation, read 17 Oct 2026
Cerebras	Partly	Published Partly — The pages do not say directly that reasoning works with tools; the raw reasoning format does not work with strict schemas. “`reasoning_format`: "raw" is incompatible with `json_object` and `json_schema` response formats.” provider documentation, read 17 Oct 2026
Mistral	Not stated	Published Not stated — The page does not say whether reasoning works together with tools; for agent tasks it recommends `reasoning_effort` “high”. “mistral-large-4-0: Supports adjustable reasoning via the `reasoning_effort` parameter.” provider documentation, read 17 Oct 2026
Z.ai (GLM)	Yes	Published Yes — “Stream Tool Call is a unique feature of Z.ai's latest model, allowing real-time access to reasoning processes, response content, and tool call information during tool invocation” provider documentation, read 17 Oct 2026

Cache covers tool definitions

The prompt cache includes the tool definitions, which matters with a long list of tools. What breaks: Where the cache does not cover tool definitions, a long tool list is paid as fresh input on every request.

Provider	Support	In the provider’s words, or in our test call
Anthropic	Yes	Published Yes — “Prompt caching references the entire prompt: tools, system, and messages (in that order), up to and including the block designated with `cache_control`.” provider documentation, read 17 Oct 2026
OpenAI	Yes	Published Yes — “OpenAI caches the model's full rendered context including OpenAI-provided instructions, developer messages, tool definitions” provider documentation, read 17 Oct 2026
Google Gemini	Not stated	Published Not stated — The page does not say whether tool definitions are part of the cached prefix. “Implicit caching Implicit caching is enabled by default for all Gemini 2.5 and newer models.” provider documentation, read 17 Oct 2026
DeepSeek	Not stated	Published Not stated — Automatic prefix cache; the page does not mention tool definitions. “The DeepSeek API Context Caching on Disk Technology is enabled by default for all users, allowing them to benefit without needing to modify their code.” provider documentation, read 17 Oct 2026
Alibaba Cloud (Qwen)	Not stated	Published Not stated — The page does not mention tool definitions. “Context Cache reduces redundant computation by caching the common prefix of these requests.” provider documentation, read 17 Oct 2026
xAI	Not stated	Unknown Not stated — We found no page on caching together with tools. xAI’s rate-limit page says cached prompt tokens still count toward its tokens-per-minute limit.
Groq	Yes	Published Yes — “Prefixes can include system prompts, tool definitions, few-shot examples, and more.” provider documentation, read 17 Oct 2026
Cerebras	Yes	Published Yes — “This includes system prompts, tool definitions, and few-shot examples.” provider documentation, read 17 Oct 2026
Mistral	Not stated	Published Not stated — Cache by `prompt_cache_key`; the page does not mention tool definitions. “Prompt caching lets you reuse previously computed prompt tokens when requests share the same prefix, meaning the same beginning of the prompt.” provider documentation, read 17 Oct 2026
Z.ai (GLM)	Not stated	Published Not stated — Implicit cache; the page does not mention tool definitions. “Context caching functionality significantly reduces token consumption and response latency by caching repeated context content.” provider documentation, read 17 Oct 2026

69 statements: 56 quoted from provider documentation read 17 Oct 2026, 11 from our own single test calls, 2 not stated anywhere we looked. A provider can change a feature without notice; a test call shows one day, not a promise. Prices for the same models: API pricing.

All API prices · Cheapest per task · Long-context cliff · Caching and batch · Rate limits by tier · Rate limit planner · Task cost calculator · All models on OpenRouter · How we check prices