benchmarks
Gemini 2.5 Pro is Google DeepMind’s reasoning model for text generation, coding and multimodal understanding, with a documented million-token input capacity. Current service availability and lifecycle status: Not yet confirmed by public reporting.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| Aider PolyglotCode editing across six programming languages | 82.2%Jul 2025 | View |
| AIMECompetition mathematics | 88%Jul 2025 | View |
| GPQA DiamondPhD-level science questions | 86.4%Jul 2025 | View |
| Humanity's Last Exam — without tools | 21.6%Jul 2025 | View |
| LiveCodeBenchFresh competitive-programming problems | 69%Jul 2025 | View |
| MMMUMultimodal (image + text) understanding | 82%Jul 2025 | View |
| SWE-bench VerifiedResolving real GitHub issues | 67.2%Jul 2025 | View |
| OTIS Mock AIME 2024-25Competition mathematics | 84.7%16 Nov 2025 | View |
| Terminal-BenchLong-horizon command-line agent tasks | 32.6%17 Jun 2025 | View |
| FrontierMath | 14.1%24 Nov 2025 | View |
| LMArena TextHuman preference Elo from blind chat votes (not a percentage) | 1458 Elo22 Oct 2026 | View |
| LMArena WebDevHuman preference Elo for building web apps (not a percentage) | 1226 Elo23 Oct 2026 | View |
Weighted mean of 10 verified benchmark results: Aider Polyglot 82.2%, AIME 88%, GPQA Diamond 86.4%, Humanity's Last Exam — without tools 21.6%, LiveCodeBench 69%, MMMU 82%, SWE-bench Verified 67.2%, OTIS Mock AIME 2024-25 84.7%, Terminal-Bench 32.6%, FrontierMath 14.1%.
Capabilities
- Built-in thinking for multistep mathematical, scientific and coding tasks.
- Understanding of text, images, audio, video and PDF inputs, with text output.
- Long-context analysis across large documents and code collections.
- Code generation, debugging and repository-level software-engineering assistance.
- Function calling and structured output for application integration.
What improved
- Makes built-in thinking central to the Pro model rather than relying on the earlier non-thinking Gemini approach.
- Improves mathematical and scientific reasoning over earlier Gemini generations in Google's published evaluations.
- Improves coding and agentic software-engineering results over earlier Gemini generations.
- Combines stronger multimodal reasoning with million-token input capacity.
Known flaws & limitations
- Can produce factual errors and unsupported conclusions; benchmark results do not establish reliability for individual answers.
- Thinking cannot be completely disabled for Gemini 2.5 Pro, limiting control over reasoning overhead.
- The standard model generates text, not native image, audio or video output.
- Documented standard Gemini API prices are USD per million tokens: $1.25 input and $10 output for prompts up to 200,000 tokens; $2.50 input and $15 output above that threshold.
- Output billing includes thinking tokens, so visible response length alone does not determine cost.
Risks, safety & security issues
- Untrusted documents and tool responses can carry prompt injections; application permissions must be enforced independently of model instructions.
- Safety safeguards do not guarantee resistance to adversarial prompting or harmful-content requests.
- Coding and reasoning capabilities can support malicious software development and other dual-use cyber activities.
- Incorrect advice or inappropriate tool actions can cause harm when applications treat model outputs as authoritative.
Good fit for
- Human-reviewed coding, debugging and software maintenance.
- Analysis and synthesis of large document collections.
- Multimodal document, image and video understanding.
- Mathematical and scientific assistance with independent checking.
Poor fit for
- Unsupervised high-stakes medical, legal or financial decisions.
- Applications requiring native image, audio or video generation.
- Strict low-latency workloads requiring reasoning to be completely disabled.
- Agents with unrestricted access to production systems, credentials or sensitive data.
References
Profile last reviewed 10/27/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.