Editorial guide
LLM Context Windows Compared: Limits and Effective Use
By TechAIOrbit · Published
An advertised context window is a request-size limit, not a promise that every token will be recalled equally well. Compare limits, output, access, cost, and task quality separately.
Last verified: 2026-09-14. Values come from current official provider documentation. Aliases, preview status, product interfaces, and model limits change; verify the exact endpoint.
Direct answer
A context window is the token budget a model can consider for one request or conversation state, but providers account for input and output differently. Maximum output is a separate cap. Product UIs may impose file, message, or plan limits below the API limit. A larger window does not prove better retrieval or reasoning over long inputs.
Inclusion methodology
This small representative table includes currently documented models with an official limit and identifiable API access path. It is not a leaderboard and does not duplicate every model profile.
| Provider | Model | Context window | Max output | Access | Verified |
|---|---|---|---|---|---|
| OpenAI | GPT-6 Astra | 1.05M tokens | 128K | Responses API | 2026-09-14 |
| Anthropic | Claude Sonnet 5 | 1M tokens | 128K | Claude API and documented cloud platforms | 2026-09-14 |
| Gemini 3.1 Pro preview | 1M input tokens | Up to 64K for Gemini 3 | Gemini API | 2026-09-14 | |
| Mistral AI | Mistral Large 3 | 256K, input plus output | Reserved inside window | Mistral API | 2026-09-14 |
Do not collapse these terms
- Context window: total capacity defined by the provider.
- Input limit: prompt, history, tools, images, and document representation supplied.
- Maximum output: generated-token cap; it may be separate or consume a combined window.
- Effective context: what your evaluated task uses reliably; this is empirical.
- API versus product limit: chat products can add upload, plan, message, or workspace restrictions.
What 128K, 256K, and 1M mean
Conversion varies by language, formatting, code, and tokenizer. A rough English planning range is 0.6–0.8 words per token: 128K may represent about 77,000–102,000 words, 256K about 154,000–205,000, and 1M about 600,000–800,000. Count with the target provider before sending.
Why the biggest window may not be best
Long prompts increase latency and uncached input cost. Relevant evidence can be diluted by distractors; quality depends on placement, structure, model behavior, and the question. Evaluate known evidence at different positions and lengths. Prompt caching may reduce repeated-input cost, but rules and charges differ.
When RAG is preferable
Use retrieval-augmented generation when the corpus changes often, access control differs by document, citations must map to narrow passages, or sending everything is wasteful. Use direct long context when complete relationships matter and cost fits. Hybrid systems retrieve first, then place bounded evidence into context.
Evaluation checklist
- Pin the exact model ID and API.
- Count with that provider's tokenizer.
- Reserve output and tool-result capacity.
- Test known facts at early, middle, and late positions.
- Measure accuracy, latency, cached/uncached cost, and truncation.
- Recheck product UI limits separately.
Open model profiles, estimate spend with the API cost calculator, use Compare for two-model facts, and see the local LLM guide for on-device constraints.
Frequently asked questions
- Does a larger context window mean a better LLM?
- No. It means the request can accept more tokens. Retrieval quality, reasoning, latency, output limits, and cost need separate evaluation.
- Is the context window input only?
- It depends on the provider and model. Some publish separate input and output limits; others require input and output together to fit.