AI context window leaderboard

The 1M-token tier is now the default

In 2024, a 200K token context window was considered large. By June 2026, every Anthropic Claude model, both DeepSeek V4 models, Grok 4.3, Kimi K2.6, and Gemini 3.5 Flash all support at least 1 million tokens. The 1M tier went from a premium feature to the new baseline in under two years.

GPT-5 is the notable exception among flagship models — its 400K limit is about 2.5× smaller than the new standard. For workloads that don't need more than 400K tokens, this doesn't matter. For those that do, it's a functional constraint.

The 10M outlier: Llama 4 Scout

Llama 4 Scout's 10 million token context is an engineering milestone. It's a self-hosted open-weight model using a MoE architecture with 109 billion total but only 17 billion active parameters per token. The 10M window enables use cases that no other model supports: full large codebase analysis, entire book ingestion, months of conversation history — all in one context.

Whether retrieval quality holds across 10M tokens at the same fidelity as 200K is a different question. Longer contexts tend to produce more "lost in the middle" failures. Use the 10M capability judiciously and test retrieval from various positions before relying on it in production.

Context size vs cost

Larger context costs more per call since you pay per token. A 200K-token call costs 5× a 40K-token call at the same per-token rate. The cheapest path for long-context workloads is often chunking with a small, cheap model rather than using a 1M-token window with an expensive flagship. Only use the large context window when retrieval quality requires it.

Operational checks before buying more context

Maximum context is a capacity number, not a guarantee that the model will use every token well. Test retrieval from the beginning, middle, and end of the prompt; long-context failures often appear in the middle of large documents where the model has accepted the text but does not reliably use it.

Also measure latency and prompt construction cost. A million-token call may simplify your architecture, but it can be slower and more expensive than retrieval plus a smaller context when the relevant evidence is only a few pages. The right choice is the one that returns the correct answer with the least total context, not the largest advertised window.

What to log during long-context tests

When evaluating a large context window, log more than success or failure. Record prompt size, relevant-evidence position, latency, output length, and whether the model cited the correct part of the source. A model that succeeds only when the answer appears near the end of the prompt is not reliable enough for whole-document automation.

For codebases and legal documents, a hybrid approach often wins: retrieval selects the relevant files or clauses, then a strong model reasons over a smaller context. Use the full window when the relationships across distant sections matter, not merely because the provider accepts more tokens.

For procurement, ask providers whether the advertised context size is available on the exact endpoint, region, and plan you intend to use. Some platforms expose smaller windows through partner clouds or older aliases, and that difference can break a design built around full-document prompts.