Chat models - APIAny.AI Docs

Chat models can be exposed through:

/v1/chat/completions
/v1/responses
/v1/messages
/v1beta/models/{model}:generateContent

Pricing can include:

Input tokens.
Output tokens.
Cached input tokens.
Minimum request credits.

Context window and maximum output limits are model-level capability metadata.

Available models

Grok 4.7

Use the independent model ID grok-4.7 with /v1/chat/completions. Grok 4.6 remains a separate model. APIAny input and cached-input pricing is 1.2 credits per 1,000 tokens; output pricing is 2.4 credits per 1,000 tokens, matching Grok 4.6 at launch. Credit-pack discounts continue to apply.

{"model":"grok-4.7","messages":[{"role":"user","content":"Explain caching in three concise points."}],"max_tokens":1024}

This integration initially advertises text input and text output only. Vision, document input, context-window limits and reasoning controls are not verified. Check /v1/models for current availability and pricing. Launch testing observed refusals for some harmless exact-phrase echo requests, despite successful ordinary text and code responses. Do not rely on exact-string echo behavior as an availability probe. This is an observed model limitation, not a guarantee that every instruction will be followed.

Model ID	Description
`gpt-5.5`	OpenAI GPT-5.5 — flagship model for advanced reasoning, coding, and complex multi-step instructions.
`gpt-5.4`	OpenAI GPT-5.4 — high-capability model for advanced reasoning, coding, and agentic workflows.
`gpt-5.4-mini`	OpenAI GPT-5.4 mini — balanced speed, quality, and cost for high-volume production use.
`gpt-5.4-nano`	OpenAI GPT-5.4 nano — the fastest, lowest-cost GPT-5.4 tier for simple, high-throughput tasks.
`gpt-4o-mini`	OpenAI GPT-4o mini — fast, affordable multimodal model for everyday chat and lightweight tasks.
`claude-opus-4-8`	Anthropic Claude Opus 4.8 — frontier model for the most demanding reasoning, coding, and long-form work.
`claude-sonnet-4-6`	Anthropic Claude Sonnet 4.6 — a balanced frontier model delivering strong reasoning at production speed and cost.
`gemini-3.1-pro`	Google Gemini 3.1 Pro — top-tier multimodal model for complex reasoning, long context, and demanding production workloads.
`gemini-3.5-flash`	Google Gemini 3.5 Flash — fast, cost-efficient multimodal model with strong reasoning and a large context window.
`gemini-3.1-flash-lite`	Google Gemini 3.1 Flash Lite — fast, economical model with improved reasoning over the 2.5 generation.
`gemini-2.5-flash-lite`	Google Gemini 2.5 Flash Lite — ultra-low-cost, low-latency model for high-volume everyday tasks.
`deepseek-v4`	DeepSeek V4 — strong reasoning and coding performance at a highly competitive price.
`deepseek-v4-flash`	DeepSeek V4 Flash — a latency-optimized DeepSeek V4 variant for high-volume, low-cost inference.
`deepseek-v4.1-flash`	DeepSeek V4.1 Flash — multimodal reasoning and coding with a 1M-token context.
`glm-5.3-flash`	GLM 5.3 Flash — multimodal coding and reasoning with always-on thinking.
`glm-5.1`	Zhipu GLM-5.1 — a capable bilingual (Chinese/English) model for reasoning, coding, and agent applications.
`minimax-m3`	MiniMax M3 — a capable large language model for long-context reasoning, tool use, and multilingual chat.

Reasoning effort

DeepSeek Flash and GLM 5.3 Flash

deepseek-v4-flash and deepseek-v4.1-flash are separate models, not aliases. V4 Flash accepts text; V4.1 Flash accepts text and images. Both support up to 393,216 output tokens within a 1,048,576-token context window. GLM 5.3 Flash supports the same context window and up to 131,072 output tokens. Use /v1/models for current availability and prices. glm-5.2 is no longer available for new requests; previous usage records remain available. DeepSeek Flash accepts none, low, high, and max. Compatibility values minimal map to low; medium and xhigh map to high. In Chat Completions, thinking: { "type": "disabled" } or effort none disables thinking. Do not combine contradictory settings. For deepseek-v4.1-flash Chat Completions, thinking is enabled by default. Forced tools (tool_choice: "required" or a named function) require reasoning_effort: "none" or thinking: { "type": "disabled" }. Otherwise the request returns HTTP 400 with invalid_tool_choice. To keep thinking enabled, use tool_choice: "auto"; APIAny.AI does not silently disable thinking for you. GLM 5.3 Flash always thinks. Its effort values are low, high, and max (model default); disabling thinking returns 400. For a quick smoke test, use low. Allocate enough tokens for thinking plus the final answer, for example max_tokens: 32768. Explicit limits are honored, not silently increased. max_completion_tokens is accepted as an alias; different simultaneous limits return 400. When continuing a tool conversation with these models, preserve the assistant’s reasoning_content, tool_calls, and matching tool result messages. Cache-hit input tokens are included in total input tokens, not added again. Submit images with Chat image_url or Responses input_image. glm-5.3-flash additionally accepts documents (file / input_file) and video; documents, audio, and video are not accepted for deepseek-v4.1-flash or deepseek-v4-flash.

Official model IDs

deepseek-flash and deepseek-v4-pro are accepted as equivalent IDs for deepseek-v4.1-flash and deepseek-v4, so a client written against those official names works unchanged. Usage and billing are recorded once under the APIAny.AI model ID. GET /v1/models lists both IDs and marks the equivalent entry with alias_of.

{
  "model": "deepseek-v4.1-flash",
  "messages": [{ "role": "user", "content": "Explain how to safely retry an API request." }],
  "reasoning_effort": "low",
  "max_tokens": 32768,
  "stream": true
}

For GLM, use "model": "glm-5.3-flash" with the same example. In Responses, use input, max_output_tokens, and reasoning: { "effort": "low" }. For Responses tool follow-ups, append the complete returned output items (including reasoning and function calls) and the matching function_call_output to your next input. Do not discard reasoning items. DeepSeek Responses is stateless: resend the conversation history instead of relying on previous_response_id. Use reasoning_effort in Chat Completions and reasoning: { "effort": "high" } in Responses. Each endpoint also accepts the other form as a compatibility alias. If both are present, their values must match; conflicting values return 400. Omitting effort preserves the model default. The syntax accepts none, minimal, low, medium, high, xhigh, max, and ultra. This is not a promise that every model supports every value. Invalid values and values outside a documented model’s supported set return 400.

Model ID	Supported effort	Default
`kimi-k3`	`low`, `high`, `max`	`max`
`glm-5.3-flash`	`low`, `high`, `max`	`max`
`deepseek-v4.1-flash`, `deepseek-v4-flash`	`none`, `minimal`, `low`, `medium`, `high`, `xhigh`, `max`, `ultra`	`high`
`deepseek-v4`, `deepseek-v4-pro`	`none`, `minimal`, `low`, `medium`, `high`, `xhigh`, `max`, `ultra`	`high`
`gpt-5.6-sol`, `gpt-5.6-terra`	`none`, `low`, `medium`, `high`, `xhigh`, `max`	`medium`
`gpt-6-astra`	`low`, `medium`, `high`, `xhigh`, `max`	—
`gpt-5.5`	`none`, `low`, `medium`, `high`, `xhigh`	`medium`
`gpt-5.4`, `gpt-5.4-mini`, `gpt-5.4-nano`	`none`, `low`, `medium`, `high`, `xhigh`	`none`
`gpt-4o-mini`	Not supported	—
`claude-opus-5`, `claude-opus-4-8`	`low`, `medium`, `high`, `xhigh`, `max`	`high`
`claude-sonnet-4-6`	`low`, `medium`, `high`, `max`	`high`
`grok-4.5`, `grok-4.5-latest`, `grok-4.6`, `grok-4.7`	`low`, `medium`, `high`, `xhigh`	`high`
`gemini-3.1-pro`	`low`, `medium`, `high`	`high`
`gemini-3.1-flash-lite`	`minimal`, `low`, `medium`, `high`	`minimal`
`gemini-3.5-flash`	`minimal`, `low`, `medium`, `high`	`medium`
`gemini-2.5-flash-lite`	`none`, `minimal`, `low`, `medium`, `high`	—
`minimax-m3`	Not supported; use `thinking`	—

Reasoning is always on for kimi-k3, glm-5.3-flash, gpt-6-astra, the Grok 4.5 family and the Gemini 3 family; a value that disables reasoning returns 400. gpt-5.4 and its mini and nano variants reason only when you ask for low or higher. minimax-m3 and gpt-4o-mini take no effort parameter at all; minimax-m3 exposes the thinking object instead. Some models also accept the compatibility spellings that older integrations used: grok-4.5, grok-4.5-latest, grok-4.6 and grok-4.7 accept max (executed as xhigh) and minimal (executed as low), and gemini-3.1-pro accepts minimal (executed as low). They are normalized before the request runs, so the model always receives an official level. capability_metadata.reasoning.compatEfforts publishes these mappings; new integrations should use the values in efforts. For Anthropic Messages, use output_config.effort. Native thinking remains independent: effort does not enable thinking or select budget_tokens. For Gemini 3, native requests use generationConfig.thinkingConfig.thinkingLevel; the accepted levels vary by model. Gemini 2.5 uses thinkingBudget. In the OpenAI-compatible format, Gemini 2.5 maps minimal/low to 1024, medium to 8192, high to 24576, and none to 0 (except Pro). Do not combine explicit effort with native thinkingLevel or thinkingBudget: this returns 400 even if the settings are equivalent. includeThoughts alone can be combined with effort. For other OpenAI-compatible models, syntactically valid efforts are accepted for model execution without a declared support guarantee. Consult the model’s capabilities before relying on an effort value. Output limits such as max_completion_tokens, max_output_tokens, and the native equivalents include reasoning and final-answer tokens. Effort is not an output-token limit. A request can still exhaust its budget (finish_reason: "length" in Chat Completions) and return an incomplete or empty final answer.

Model capability metadata

GET /v1/models returns a capability_metadata object for every model that declares its capabilities. Read it to configure a model without guessing:

input lists the accepted input types: text, image, video, audio, file.
output lists the produced types.
features reports stream, tools, vision, jsonMode, and whether the model reasons at all (thinking).
reasoning describes the reasoning controls: mode (always-on, optional, or none), the accepted efforts, the defaultEffort that applies when you omit effort, and disableWith for models whose reasoning can be turned off.
limits reports contextWindow and maxOutputTokens.

reasoning.mode is the field to check before you send an effort value: always-on means reasoning cannot be disabled, optional means disableWith applies, and none means the model has no reasoning mode.

{
  "id": "glm-5.3-flash",
  "capability_metadata": {
    "input": { "text": true, "image": true, "video": true },
    "features": { "stream": true, "tools": true, "thinking": true },
    "reasoning": {
      "supported": true,
      "mode": "always-on",
      "efforts": ["low", "high", "max"],
      "defaultEffort": "max"
    },
    "limits": { "contextWindow": 1048576, "maxOutputTokens": 131072 }
  }
}

API format references: OpenAI, Anthropic Messages, Gemini, Gemini OpenAI.

​Available models

​Grok 4.7

​Reasoning effort

​DeepSeek Flash and GLM 5.3 Flash

​Official model IDs

​Model capability metadata