Context window size has become one of the most marketed specifications in AI, and also one of the most misunderstood. Here is a grounded look at what current context capabilities actually enable.
Gemini 2.5 Pro's 1 million token window — the largest available as a standard offering from a major provider — can hold approximately 750,000 words of text, or around 3,000 pages of a document. That is enough to fit most individual legal contracts, annual reports, or codebases for small-to-medium applications in a single prompt. Google has demonstrated use cases including feeding an entire feature-length film script plus its shooting schedule into one prompt to answer production questions.
Anthropic offers extended context for Claude 4 models through its API with a 2 million token ceiling, though this is not a standard feature and requires coordination with the sales team for enterprise access. The 200k window available by default handles the majority of practical use cases. At 2 million tokens, you can fit a full-length novel series, an entire legal case file including exhibits, or a software repository with 50,000+ lines of code.
The practical limitation is not the window size — it is model attention over long contexts. Both Gemini 2.5 Pro and Claude 4 Opus exhibit what researchers call "lost in the middle" degradation: information placed in the middle of a very long context is retrieved and applied less reliably than information near the beginning or end. Google's internal testing shows retrieval accuracy above 99 percent for documents under 200k tokens and declining meaningfully above 500k. Anthropic's published research shows similar patterns.
This means the 1M+ context window is most valuable when the relevant information is densely distributed throughout the document rather than concentrated in the middle — or when the task involves summary and synthesis rather than precise fact retrieval.
The practical winners from large context windows are legal tech, financial analysis, and enterprise code review. For most consumer-facing AI applications, a 128k window is more than sufficient, and the cost difference between providers matters more than the ceiling.