Context Window Sizes: Every Major LLM Compared for 2026

Mid-2026 Context Window Comparison

Prices reflect standard API rates as of July 2026 and change frequently. Self-hosted model costs depend on hardware and hosting provider. Some models have long-context surcharges: GPT-5.5 charges 2x input pricing for prompts over 272K tokens, and Gemini 2.5 Pro charges $2.50 per million input tokens (instead of $1.25) for prompts over 200K tokens. Claude Opus 4.6 and Sonnet 4.6 charge standard rates across the full 1M window with no long-context surcharge.

What These Token Counts Mean in Practice

Token counts are abstract until you translate them into content you can visualize. One token is roughly 0.75 words in English, or about 4 characters. Here is what various window sizes hold in practical terms:

128,000 tokens (~200 pages): About the length of a short novel. This fits a substantial system prompt (1,000-2,000 tokens), a long conversation history (20-30 turns), and 10-15 retrieved document chunks for RAG. For the majority of production applications, 128K is more than sufficient. GPT-4o and GPT-4o mini operate at this tier and handle most workloads without hitting the ceiling.

200,000 tokens (~300 pages): Claude Haiku 4.5 sits here. Enough to process a technical book, a full codebase of a small project, or several long documents simultaneously. The extra headroom over 128K matters when you are combining long system prompts with large retrieved contexts and multi-turn conversation history.

400,000 tokens (~600 pages): GPT-5 provides this midpoint tier. Fits multiple long documents, extensive conversation histories, and detailed system instructions all at once. Practical for applications that need to reference several complete documents in a single request without chunking.

1,000,000 tokens (~1,500 pages): The new frontier standard. Claude Opus 4.6, Claude Sonnet 4.6, Gemini 2.5 Pro, GPT-5.5, and DeepSeek V4 all offer windows at this scale. A million tokens holds multiple textbooks, an entire codebase of a mid-sized application, or months of conversation history. The question at this scale is not whether your content fits, but whether the model can reliably use all of it.

10,000,000 tokens (~15,000 pages): Llama 4 Scout's headline figure. Independent benchmarking (RULER tests) suggests effective performance degrades significantly beyond 5-6 million tokens, making the practical ceiling closer to 5M for most tasks. At this scale, you could theoretically load an entire enterprise documentation library, but attention quality across 15,000 pages of text is an active research problem, not a solved one.

The Cost of Filling a Context Window

Context window size and cost are directly connected because every input token requires computation. The spread in cost for filling the same window size is enormous across providers. Here is what it costs to fill common context budgets with input tokens alone (output costs add on top):

Filling a full 1M-token window costs $0.09 on DeepSeek V4 Flash and $5.00 on GPT-5.5, a 55x spread. For an application making 1,000 calls per day at 100K tokens average input, the monthly cost difference between DeepSeek V4 Flash ($270) and GPT-5.5 ($15,000) is the difference between a rounding error and a serious infrastructure line item. This is why context management matters regardless of window size: every token you can remove from the context directly reduces your bill.

Prompt caching changes the math significantly. Anthropic, OpenAI, and Google all offer caching that reduces input costs by 75-90% for repeated prompt prefixes (system prompts, static document context). If your application sends the same 50K-token system prompt on every call, caching that prefix means you pay full price once, then 10-25% of the price on subsequent calls. For high-volume applications with stable prompt prefixes, caching effectively eliminates the cost penalty of large contexts.

Effective vs Advertised Window Size

The advertised context window is a theoretical maximum. The effective window, where the model reliably attends to and uses the information you provide, is smaller. Several well-documented phenomena reduce the effective window:

Lost in the middle: Research consistently shows that LLMs pay more attention to information at the beginning and end of the context than to information in the middle. For contexts over 20,000 tokens, this effect becomes measurable. A key fact placed at position 60,000 in a 100,000-token context has a 20-30% lower chance of being correctly used compared to the same fact placed in the first 5,000 tokens. This effect has improved in newer models (Claude Opus 4.6 and GPT-5.5 handle it better than their predecessors) but has not been eliminated.

Attention dilution: As context length increases, each individual token receives proportionally less of the model's attention budget. A relevant 500-token passage competes for attention with every other token. In a 10K-token context, it gets roughly 5% of the attention. In a 1M-token context, it gets 0.05%. This mathematical reality means that the signal-to-noise ratio of your context matters more at larger scales. A carefully curated 50K-token context with only relevant information will outperform a 500K-token context where 90% of the content is tangentially related.

Instruction following degradation: Long contexts can weaken the model's adherence to system prompt instructions. The system prompt sits at the beginning of the context, and as the distance between the system prompt and the current query grows, the system prompt's influence on the response weakens. Models with instruction hierarchy features (Claude's system prompt anchoring, for example) mitigate this, but the effect is still measurable at 500K+ token contexts.

Practical guideline: Plan around 40-50% of the advertised window as your effective size for tasks requiring precise retrieval and reasoning. For a 128K model, target 50-60K usable tokens. For a 1M model, target 400-500K. You can send more, and the model will process it, but response quality will be measurably better with curated context than with maximum context. The exception is tasks where the model only needs to find a specific piece of information (needle-in-a-haystack), where frontier models now perform well across the full window.

Model Selection by Use Case

High-volume chatbots and simple Q&A: GPT-4o mini, Claude Haiku 4.5, DeepSeek V4 Flash, or Gemini 2.5 Flash provide the best price-to-performance ratio. Their windows (128K-1M) exceed what chatbot applications need, and their per-token costs make millions of daily calls affordable. Use a 10K-15K token working context with a sliding window or summarization for conversation history.

RAG applications: Claude Sonnet 4.6, GPT-5, or Gemini 2.5 Pro offer the reasoning quality needed for complex retrieval tasks at reasonable cost. Use 20K-40K tokens of curated retrieved context rather than filling the window. The quality improvement from better retrieval (fewer, more relevant chunks) consistently outperforms the quality improvement from larger context (more chunks of mixed relevance). Enable prompt caching for the system prompt to reduce costs on repeated calls.

Full document processing: For tasks that genuinely require processing an entire long document in a single pass (legal contract review, full codebase analysis, book-length summarization), the 1M-token models are the correct choice. Gemini 2.5 Pro and DeepSeek V4 Pro offer the best cost per token at this scale. Consider whether chunking with map-reduce can achieve the same result at lower cost before committing to a full-window pass, because a single 1M-token call to GPT-5.5 costs $5.00 in input alone.

Complex reasoning and analysis: Claude Opus 4.6 and GPT-5.5 provide the highest reasoning quality. Their 1M+ windows and strong attention characteristics make them suitable for tasks where accuracy matters more than cost. Reserve these models for high-value tasks (legal analysis, medical summarization, complex code review) rather than routine queries.

Self-hosted and open-source: Llama 4 Scout offers the largest window (10M tokens) for teams that can self-host. Llama 3.3 at 128K remains the most battle-tested open-source option. Self-hosting eliminates per-token API costs but introduces infrastructure costs (GPU servers, memory for KV cache at long contexts) that can exceed API costs at low volume. Self-hosting makes economic sense above roughly 100K calls per day, where the fixed infrastructure cost is amortized across enough requests.

The External Memory Alternative

For applications where the knowledge base exceeds any context window, or where you want to avoid the cost and quality tradeoffs of very long contexts, external memory is the better architecture. Instead of choosing a bigger model with a bigger window, store knowledge in a persistent memory or retrieval system and fetch only the relevant subset for each query. This approach works with any model size, keeps costs proportional to query complexity rather than knowledge base size, and avoids the attention degradation that comes with very long contexts.

A 10K-token curated context drawn from a 100-million-token knowledge base via intelligent retrieval will consistently outperform a 1M-token context dump of the same knowledge base. The retrieval step acts as a relevance filter that the attention mechanism alone cannot replicate at scale. For most production applications, the right architecture is a modest context window (50K-100K tokens) combined with a strong retrieval layer, not a maximum-size context window filled with everything available.