Gemini 2.5 Pro Reasoning: pricing, context window and benchmarks
Gemini 2.5 Pro Reasoning is a frontier model from Google DeepMind. It offers a 128,000-token context window, costs $16.81 per million input tokens and $67.95 per million output tokens, and averages 83.8 across the index's benchmark suite.
Pricing
Gemini 2.5 Pro Reasoning costs $16.81 per million input tokens and $67.95 per million output tokens. A blended 3:1 input-to-output mix works out to about $29.59 per million tokens. Cached input is listed at $2.20 per million tokens.
| Item | Price | Notes |
|---|---|---|
| Input | $16.81 | per 1M tokens |
| Output | $67.95 | per 1M tokens |
| Cached input | $2.20 | per 1M tokens |
| Blended 3:1 | $29.59 | estimate per 1M tokens |
| 1M in + 1M out | $84.76 | estimate |
Benchmarks
Scores below are modelled estimates on a 0–100 scale. Use them to compare relative position, not as official vendor results.
| Benchmark | Score | Rank | Percentile | Relative |
|---|---|---|---|---|
| MMLU-Pro | 86.8 | 40 | 93th | |
| GPQA Diamond | 80.0 | 56 | 91th | |
| SWE-bench Verified | 71.2 | 93 | 85th | |
| HumanEval | 86.7 | 94 | 84th | |
| MATH-500 | 87.9 | 68 | 89th | |
| AIME 2025 | 84.2 | 47 | 92th | |
| LiveCodeBench | 74.8 | 98 | 84th | |
| IFEval | 90.8 | 59 | 90th | |
| TruthfulQA | 84.0 | 83 | 86th | |
| HellaSwag | 92.1 | 42 | 93th | |
| MMMU | 82.8 | 25 | 96th | |
| Arena Elo | 1472.1 | 30 | 95th |
What Gemini 2.5 Pro Reasoning is good for
Gemini 2.5 Pro Reasoning sits in the frontier tier of the index. It is aimed at work where answer quality matters more than unit cost: hard reasoning, multi-step coding, document analysis and multimodal inputs. At $16.81 per million input tokens and a 128,000-token context window, it is a premium option, so reserve it for tasks where a cheaper model measurably fails.
Strengths
- Top-tier average benchmark score (83.8), placing it #62 of 600.
- Large 128,000-token context window for long documents and agent traces.
- Accepts more than text, including image input where the provider supports it.
- Reasoning-oriented profile, useful for multi-step problems and structured analysis.
- Coding benchmarks are a relative strength within its tier.
Trade-offs
- Proprietary licence: you cannot download the weights or self-host the model.
- Premium input pricing at $16.81 per million tokens.
- Cost can climb quickly on long outputs and high-volume traffic.
How it compares
| Model | Provider | Context | Input / 1M | Avg score |
|---|---|---|---|---|
| Claude Haiku 4.5 Max | Anthropic | 200000 | $15.98 | 83.8 |
| Claude Sonnet 4.5 Vision | Anthropic | 128000 | $16.99 | 83.8 |
| GPT-4.1 Reasoning | OpenAI | 1000000 | $16.97 | 83.8 |
| Gemini 2.5 Pro Reasoning | Google DeepMind | 128000 | $16.81 | 83.8 |
| Grok 3 Pro | xAI | 400000 | $16.61 | 83.8 |
| DeepSeek-V3.2 Instruct | DeepSeek | 128000 | $5.91 | 83.7 |
FAQ
How much does Gemini 2.5 Pro Reasoning cost?
Gemini 2.5 Pro Reasoning is listed at $16.81 per million input tokens and $67.95 per million output tokens. Cached input is $2.20 per million tokens. A blended 3:1 input-to-output mix is about $29.59 per million tokens.
What is the context window of Gemini 2.5 Pro Reasoning?
Gemini 2.5 Pro Reasoning has a context window of 128,000 tokens and can emit up to 4,096 output tokens per request in this index.
Is Gemini 2.5 Pro Reasoning open weights?
No. Gemini 2.5 Pro Reasoning is a proprietary model with a proprietary licence, so you cannot download the weights.
How does Gemini 2.5 Pro Reasoning score on benchmarks?
Gemini 2.5 Pro Reasoning averages 83.8 across the scored benchmarks in this index, ranking #62 of 600 models. Its strongest relative result is Arena Elo at 1472.1.
What is Gemini 2.5 Pro Reasoning best for?
Gemini 2.5 Pro Reasoning sits in the frontier tier of the index. It is aimed at work where answer quality matters more than unit cost: hard reasoning, multi-step coding, document analysis and multimodal inputs. At $16.81 per million input tokens and a 128,000-token context window, it is a premium option, so reserve it for tasks where a cheaper model measurably fails.