Context windows in practice

What a context window actually is

A context window is the maximum number of tokens a single request can carry as input. It is the union of the system prompt, the conversation history, any retrieved documents, tool definitions, and the new user message. A model with a 1,000,000-token context window can accept up to that many tokens in one inference call — not across a session, not across an agent loop, but in one shot.

The catalogue records context window as a verified field on the model entity. Where it is not directly published in the vendor's docs, the record renders the canonical unverified-data label rather than guessing.

Input limit vs output limit

Context window is an input limit. The number of tokens a model can generate in a single response is a separate field — max output tokens — and is usually a much smaller number than the context window. Claude Opus 4.7 advertises a 1M-token context window and a 128k-token max output (300k via the Message Batches beta header). Gemini 2.5 Pro advertises a ~1M-token context and a 65,536-token max output. Mixing these up is a common source of cost and design errors: a workload that needs to read a long document is different from one that needs to write one.

See /research/model-output-limits for the output side.

Tokenizer differences across providers

One million tokens at provider A is not the same payload as one million tokens at provider B. Each provider ships its own tokenizer, and a given fixed text fragments differently depending on which tokenizer encodes it. Anthropic explicitly documents that Opus 4.7 can use up to 35% more tokens than earlier Claude generations for the same input — a fact captured as a verified note on that model's record.

The catalogue does not publish a tokenizer-normalised context window. The headline figures on the model pages are the vendor's own published numbers; a cost projection that needs precision should sample real prompts on each vendor's tokenizer.

Verified context + output matrix

DeepSeek V4 Pro's max output is not separately published on the Models & Pricing page; the field renders as the canonical unverified-data label on its model page.

Verified examples in the catalogue

The four million-token-class records currently in the catalogue are listed under Verified today above. The contrast between Claude Opus 4 (200k, deprecated) and Claude Opus 4.7 (1M, current) is a useful illustration of how quickly the context-window axis moves within a single provider generation.

The two-sided verified comparison Gemini 2.5 Pro vs Claude Opus 4.7 lays the context-window field side by side along with pricing and modality.

Long-context cost considerations

Long-context requests are sometimes cheaper per token than short ones (with batch APIs), sometimes more expensive (with prompt-size tiers), and sometimes the same. Google's Gemini API charges a surcharge for prompts >200k tokens — captured as the 1M input tokens (>200k context) unit on the pricing hub. Anthropic and DeepSeek currently have a flat per-token rate independent of prompt size.

Provider-specific cache pricing also interacts with long context — see /research/api-pricing-methodology.

What we do not claim

We do not publish a quality-of-recall metric across long contexts (the "needle in a haystack" style of evaluation lives elsewhere and warrants independent replication, not republication). We do not claim a million-token context is equivalent to a million-token context on another provider. We do not publish a latency number for long-context inference — request latency is workload-dependent and we have not yet wired an independent measurement harness.