Documented model context limits — 2026-10-12

These are public provider specifications, not live account tests or native-agent compaction triggers. 1M means one million tokens; 10M means ten million. No universal 1M or 10M default is assumed. Exact API/deployment IDs matter; older snapshots, previews and account plans can differ.

Provider Model ID / family Documented window or input limit Official source
OpenAI gpt-6.1-sol / gpt-6-astra / gpt-6-luna 1,050,000 Comparison
OpenAI gpt-5.1 (deprecated) / gpt-4o 400,000 / 128,000 5.1, 4o
Anthropic claude-opus-5-5 / sonnet-5-5 / haiku-5-5 1M Model comparison
Google gemini-3.8-flash / gemini-3.1-pro-preview 1,048,576 input Flash, Pro Preview
xAI grok-4.7 500,000 Developer guide
DeepSeek deepseek-flash / deepseek-v4-pro 1,048,576 Model metadata
Z.AI glm-5.2 / glm-5.3 1M 5.2, 5.3
Kimi Platform kimi-k3 1M Quickstart
Kimi Code k3 262,144 or 1,048,576 by plan Code model settings
MiniMax MiniMax-M3 / MiniMax-M2.7 1,000,000 / 204,800 API table
Qwen qwen3.8-max 1,000,000 Model-specific table
Doubao seed-2-1-pro/lite-260915; earlier pro-260628 1024k; earlier 256k API model list
Hunyuan TokenHub hy3 / hy4-preview 256k / 1M Model list
Baidu Qianfan ernie-5.1 / ernie-5.0 128k API model list

Gemini separates maximum input and output; most other rows describe a context window. The application must reserve instructions, tools and generation space according to the provider's semantics. Marketing K/M labels are retained in the evidence; the registry's lower decimal conversions are explicitly marked as conservative, not exact official integers. Kimi Code defaults to the smaller verified plan capacity unless a confirmed deployment override is supplied.

Google's reported 10M research demonstrations are not a current universal API allowance. The model maximum also does not prove that a Code Agent has not already compacted: its configured window or trigger can be smaller.

Machine-readable snapshot · Request policy / 中文