Thinking budget
Control how much 2.5 Flash reasons to balance latency and cost.
Our powerful and most efficient workhorse model designed for speed and low-cost.
Control how much 2.5 Flash reasons to balance latency and cost.
Understands input across text, audio, images and video.
Explore vast datasets with a 1-million token context window.
Converse in more expressive ways with native audio outputs that capture the subtle nuances of how we speak. Seamlessly switch between 24 languages, all with the same voice.
Remarkable quality, more appropriate expressivity, and prosody, delivered with low latency so you can converse fluidly.
Use natural language prompts to adapt the delivery within the conversation, steer it to adopt accents and produce a range of tones and expressions.
Gemini 2.5 can use tools and function calling during dialog allowing it to incorporate real-time information or use custom developer-built tools.
Our system is trained to discern and disregard background speech, ambient conversations and other irrelevant audio.
| Benchmark | Notes | Gemini 2.5 Flash Thinking | Gemini 2.0 Flash | OpenAI o4-mini | Claude Sonnet 3.7 64k Extended thinking | Grok 3 Beta Extended thinking | DeepSeek R1 |
|---|---|---|---|---|---|---|---|
| Input price | $/1M tokens | $0.30 | $0.10 | $1.10 | $3.00 | $3.00 | $0.55 |
| Output price | $/1M tokens | $2.50 | $0.40 | $4.40 | $15.00 | $15.00 | $2.19 |
| Reasoning & knowledge Humanity's Last Exam (no tools) | 11.0% | 5.1% | 14.3% | 8.9% | — | 8.6%* | |
| Science GPQA diamond | single attempt (pass@1) | 82.8% | 60.1% | 81.4% | 78.2% | 80.2% | 71.5% |
| multiple attempts | — | — | — | 84.8% | 84.6% | — | |
| Mathematics AIME 2025 | single attempt (pass@1) | 72.0% | 27.5% | 92.7% | 49.5% | 77.3% | 70.0% |
| multiple attempts | — | — | — | — | 93.3% | — | |
| Code generation LiveCodeBench | single attempt (pass@1) | 63.9% | 34.5% | — | — | 70.6% | 64.3% |
| Code editing Aider Polyglot | 61.9% / 56.7% whole / diff-fenced | 22.2% whole | 68.9% / 58.2% whole / diff | 64.9% diff | 53.3% diff | 56.9% diff | |
| Agentic coding SWE-bench Verified | 60.4% | — | 68.1% | 70.3% | — | 49.2% | |
| Factuality SimpleQA | 26.9% | 29.9% | — | — | 43.6% | 30.1% | |
| Factuality FACTS grounding | 85.3% | 84.6% | 62.1% | 78.8% | 74.8% | 56.8% | |
| Visual reasoning MMMU | single attempt (pass@1) | 79.7% | 71.7% | 81.6% | 75.0% | 76.0% | no MM support |
| multiple attempts | — | — | — | — | 78.0% | no MM support | |
| Image understanding Vibe-Eval (Reka) | 65.4% | 56.4% | — | — | — | no MM support | |
| Long context MRCR v2 | 128k (average) | 74.0% | 36.0% | 49.0% | — | 54.0% | 45.0% |
| 1M (pointwise) | 32.0% | 6.0% | — | — | — | — | |
| Multilingual performance Global MMLU (Lite) | 88.4% | 83.4% | — | — | — | — |
Methodology
Gemini results: All Gemini scores are pass @1 (no majority voting or parallel test time compute unless indicated otherwise). They are all run with the AI Studio API for the model-id gemini-2.5-flash-preview-05-20 and gemini-2.0-flash with default sampling settings. To reduce variance, we average over multiple trials for smaller benchmarks. Vibe-Eval results are reported using Gemini as a judge.
Non-Gemini results: All the results for non-Gemini models are sourced from providers' self reported numbers unless mentioned otherwise below. All SWE-bench Verified numbers follow official provider reports, using different scaffoldings and infrastructure. Google's scaffolding includes drawing multiple trajectories and re-scoring them using model's own judgement.
Thinking vs not-thinking: For Claude 3.7 Sonnet: GPQA, AIME 2024, MMMU come with 64k extended thinking, Aider with 32k, and HLE with 16k. Remaining results come from the non thinking model due to result availability. For Grok-3 all results come with extended reasoning except for SimpleQA (based on xAI reports) and Aider.
Single attempt vs multiple attempts: When two numbers are reported for the same eval higher number uses majority voting with n=64 for Grok models and internal scoring with parallel test time compute for Anthropic models.
Result sources: Where provider numbers are not available we report numbers from leaderboards reporting results on these benchmarks: Humanity's Last Exam results are sourced from https://agi.safe.ai/ and https://scale.com/leaderboard/humanitys_last_exam, AIME 2025 numbers are sourced from https://matharena.ai/. LiveCodeBench results are from https://livecodebench.github.io/leaderboard.html (10/1/2024 - 2/1/2025 in the UI), Aider Polyglot numbers come from https://aider.chat/docs/leaderboards/. FACTS come from https://www.kaggle.com/benchmarks/google/facts-grounding. For MRCR v2 which is not publically available yet we include 128k results as a cumulative score to ensure they can be comparable with previous results and a pointwise value for 1M context window to show the capability of the model at full length.
API costs are sourced from providers' website and are current as of May 20th.
* indicates evaluated on text problems only (without images)
Input and output price reflects text, image and video modalities.