Context Window Per Dollar Index 2026
DeepSeek V4 Flash Latest gives you about 109 million context tokens per dollar of input spend. GPT-4 gives you 273. We pulled all 434 models on OpenRouter, divided each model's context window by its input price, and ranked the results.
The ratio answers a narrow question: how much prompt can you feed a model per dollar? If your workload is input-heavy (RAG, long-document analysis, big system prompts), it is the single most useful price metric. Output pricing is a separate problem. See the OpenRouter pricing calculator for blended estimates.
Methodology
- Dataset: 434 model variants listed on OpenRouter, September 2026. Fields used: context window (tokens), input price (USD per 1M tokens).
- Ratio = ctx tokens ÷ input price per $1. The main table shows ctx-tokens per $1,000 input budget = ctx × 1000 ÷ in.
- Median ratio across all 434 models: 638,260 ctx-tokens per $1 (about 638M per $1,000).
Caveats: input price only. Output tokens usually cost 3–10x more and are excluded here. Context windows are as advertised by the provider; effective usable context can be smaller due to degradation past ~128K tokens on many models. Batch-tier variants (:batch, 50% discount) and “latest” aliases appear as separate rows, which is why some models show up twice. Prices change frequently; verify current rates before committing.
Top 20 by ctx-tokens per $1,000 input budget
| # | Model | Context | Input $/1M | Ctx-tokens per $1,000 |
|---|---|---|---|---|
| 1 | ~deepseek/deepseek-v4-flash-latest | 1.31M | $0.012 | 109,226,666,667 |
| 2 | deepseek/deepseek-v4-flash-0731 | 1.31M | $0.018 | 72,817,777,778 |
| 3 | ~z-ai/glm-flash-latest | 1.31M | $0.02 | 65,536,000,000 |
| 4 | ~deepseek/deepseek-flash-latest | 1.05M | $0.02 | 52,428,800,000 |
| 5 | qwen/qwen3.7-flash | 1.00M | $0.03 | 33,333,333,333 |
| 6 | qwen/qwen3.8-27b | 1.00M | $0.0307 | 32,573,289,902 |
| 7 | openai/gpt-6-luna-pro:batch | 1.05M | $0.05 | 21,000,000,000 |
| 8 | openai/gpt-6-luna:batch | 1.05M | $0.05 | 21,000,000,000 |
| 9 | google/gemini-2.5-flash-lite:batch | 1.05M | $0.05 | 20,971,520,000 |
| 10 | openai/gpt-4.1-nano:batch | 1.05M | $0.05 | 20,951,520,000 |
| 11 | z-ai/glm-5.3-flash:batch | 1.05M | $0.06 | 17,476,266,667 |
| 12 | openai/gpt-5-nano:batch | 400K | $0.025 | 16,000,000,000 |
| 13 | qwen/qwen3.5-flash-02-23 | 1.00M | $0.065 | 15,384,615,385 |
| 14 | deepseek/deepseek-v4-flash | 1.05M | $0.0763 | 13,742,804,718 |
| 15 | meta-llama/llama-4-scout | 1.31M | $0.1 | 13,107,200,000 |
| 16 | inclusionai/ling-3.0-flash-vl | 262K | $0.021 | 12,483,047,619 |
| 17 | inclusionai/ling-3.0-flash | 262K | $0.021 | 12,483,047,619 |
| 18 | poolside/laguna-s-2.1 | 1.05M | $0.09 | 11,650,844,444 |
| 19 | openai/gpt-6-luna-pro | 1.05M | $0.1 | 10,500,000,000 |
| 20 | openai/gpt-6-luna | 1.05M | $0.1 | 10,500,000,000 |
Bottom 5 (worst value)
| # | Model | Context | Input $/1M | Ctx-tokens per $1,000 |
|---|---|---|---|---|
| 430 | openai/gpt-3.5-turbo-16k | 16K | $3 | 5,462 |
| 431 | openai/gpt-3.5-turbo-0613 | 4K | $1 | 4,095 |
| 432 | openai/gpt-3.5-turbo-instruct | 4K | $1.5 | 2,730 |
| 433 | openai/o1-pro | 200K | $150 | 1,333 |
| 434 | openai/gpt-4 | 8K | $30 | 273 |
What stands out
The leaders are all budget-tier Chinese open-weight models with 1M+ contexts priced under $0.03/M. DeepSeek V4 Flash Latest at $0.012/M with a 1.31M window is roughly 800x better value than GPT-4, which still sells at legacy pricing. GLM Flash Latest and the Qwen flash models cluster in the 33–65M range.
The losers are legacy. GPT-4, GPT-3.5 variants, and o1-pro earn their spots through small windows, $30–$150/M pricing, or both. Nobody should be buying these in 2026 except for compatibility.
Batch tiers punch above their weight. GPT-6 Luna at $0.05/M via batch delivers 21B ctx-tokens per $1,000, double the standard tier at $0.10/M. If your workload tolerates 24-hour turnaround, batch pricing halves the effective cost.