Guides

Long-context document analysis (400K tokens)

Long prompts change the price picture: some providers bill the whole request at a higher rate past a threshold, and caching the document turns follow-up questions from expensive to cheap. Prices verified 2026-10-18.

The short answer

• Cheapest flagship model: Mistral Large 4 at $0.57 (temporary sale price; regular $1.36 / $4.18 per 1M)
• Cheapest mid-tier model: Gemini 3.8 Flash at $0.65 (price rises 2027-01-01)
• Cheapest small model: DeepSeek V4.1 Flash at $0.17

Long-context surcharges

These models bill a 400,200-token prompt at their long-context rate for the whole request:

• GPT-6 Luna: above 272K, input $0.10 → $0.20 and output $0.50 → $0.75 per 1M
• GPT-5.6 Luna: above 272K, input $0.20 → $0.40 and output $1.20 → $1.80 per 1M
• Claude Haiku 5.5: above 100K, input $0.10 → $0.50 and output $0.50 → $2.50 per 1M
• Grok 4.20 (Reasoning): above 200K, input $1.25 → $2.50 and output $2.50 → $5.00 per 1M
• Grok 4.3: above 200K, input $1.25 → $2.50 and output $2.50 → $5.00 per 1M
• GPT-6.1 Sol: above 272K, input $2.00 → $4.00 and output $10.00 → $15.00 per 1M
• Gemini 3.1 Pro (Preview): above 200K, input $2.00 → $4.00 and output $12.00 → $18.00 per 1M
• GPT-6 Sol: above 272K, input $2.00 → $4.00 and output $10.00 → $15.00 per 1M
• GPT-5.6 Terra: above 272K, input $2.00 → $4.00 and output $12.00 → $18.00 per 1M
• Grok 4.7: above 200K, input $2.00 → $4.00 and output $6.00 → $12.00 per 1M
• GPT-5.6 Sol: above 272K, input $4.00 → $8.00 and output $20.00 → $30.00 per 1M
• GPT-5.5: above 272K, input $5.00 → $10.00 and output $30.00 → $45.00 per 1M
• GPT-6 Astra: above 272K, input $10.00 → $20.00 and output $50.00 → $75.00 per 1M

16 of the 29 ranked models publish no long-context tier, so the 400K prompt is billed at their normal rate.

Caching the document

Without caching, every follow-up question pays for all 400,000 document tokens again. With caching, 10 questions on DeepSeek V4.1 Flash cost $0.17 instead of $1.22. How caching is applied differs: some providers cache automatically and do not guarantee a hit, and Google charges an hourly storage fee for explicit caches, which is not included here. Check each model page's billing notes.

Models that do not fit

GPT-5.4 mini (400K), Grok Build 0.1 (256K), Mistral Medium 3.5 (256K), Mistral Small 4 (256K), Ministral 3 8B (256K), Claude Haiku 4.5 (200K), Codestral (128K) have context windows too small for this prompt plus answer.

Trade-offs

Fitting a document in the window is not the same as using it well. Recall of details deep inside very long prompts varies by model; for repeated questions over a large, stable corpus, retrieval (RAG) can be both cheaper and more accurate. Test with questions whose answers sit in the middle of the document.

Full ranking (29 models)

USD, standard list prices. Tags mark announced price changes and temporary sale prices.

Price is not quality

This page ranks by cost only. Models differ in accuracy, instruction following, tool use, latency, rate limits and data policies, and those differences often matter more than the price gap. Shortlist two or three models from the table, run them on a sample of your real inputs, and compare cost per successful result.

Related comparisons

• GPT-6 Luna vs Gemini 3.5 Flash-Lite
• GPT-6 Luna vs DeepSeek V4.1 Flash
• Gemini 3.5 Flash-Lite vs DeepSeek V4.1 Flash
• GPT-6 Luna vs Gemini 3.1 Flash-Lite

More guides

• Cheapest LLM for coding agents
• Cheapest LLM for RAG chatbots
• Cheapest LLM for batch classification and extraction
• Cheapest LLM for high-volume customer support
• Prompt caching explained, with savings

FAQ