Every model’s context window, in one place

Effective vs. nominal context. Several long-context models advertise large windows but degrade past a certain length on real-world tasks — this page records the documented nominal capacity, not benchmark-measured effective length. The latter is benchmark-dependent and out of scope per the section’s no-benchmarks rule.

Input vs. output limits. Some providers document a single “context window” that includes both prompt and response tokens; others publish a separate max-output cap alongside it. Which is which varies by lab, by generation, and by whether the model is hosted or open-weights — a lab that documents one number for its open checkpoints often documents two for the hosted SKU built on them. Where a separate cap is documented, the Output column shows it; where it isn’t, the column reads — and the row treats the input window as the whole budget.

Beta and tier-gated context. Some providers ship a default context size for the standard API and a larger one behind a beta flag, batch endpoint, or paid tier. The headline number on this page is the standard-API value documented as generally available; the per-row notes call out when a beta or tier-gated extended window exists.

Open-weights inference. For open-weights models (Llama, DeepSeek, Mistral, Qwen) the “context window” is the value the model card claims; serving infrastructure (vLLM, Together, Fireworks, Hugging Face Inference) often caps the deployed window lower for memory reasons. Always check the specific endpoint’s docs before relying on the full nominal window.

Tokenizer differences. One token is not a fixed unit across providers. OpenAI’s o200k tokenizer, Anthropic’s tokenizer, Google’s SentencePiece, and Meta’s tiktoken-derived tokenizers all produce different token counts for identical text. Compare context windows in tokens, not in characters or pages, but treat them as a same-provider apples-to-apples comparison rather than a strict cross-provider one.