benchmarks
Gemini 2.5 Flash is Google DeepMind's multimodal, text-generating model with configurable reasoning, a million-token input window and support for tool calling. Its documented strengths suit high-volume analysis and assistant workflows, but factual accuracy and safe handling of untrusted content still require application-level controls.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| AIMECompetition mathematics | 72%Jul 2025 | View |
| GPQA DiamondPhD-level science questions | 82.8%Jul 2025 | View |
| MMMUMultimodal (image + text) understanding | 79.7%Jul 2025 | View |
| Terminal-BenchLong-horizon command-line agent tasks | 17.1%17 Jun 2025 | View |
| FrontierMath | 4.8%18 Dec 2025 | View |
| LMArena TextHuman preference Elo from blind chat votes (not a percentage) | 1417 Elo22 Oct 2026 | View |
Weighted mean of 5 verified benchmark results: AIME 72%, GPQA Diamond 82.8%, MMMU 79.7%, Terminal-Bench 17.1%, FrontierMath 4.8%.
Capabilities
- Processes text, images, video and audio as inputs and generates text output.
- Supports configurable thinking budgets to balance reasoning effort, latency and cost.
- Accepts up to 1,048,576 input tokens for long-document and multimodal analysis.
- Supports function calling and structured output for application integration.
- Supports Google Search grounding and code execution through supported API tools.
What improved
- Adds configurable reasoning to the Flash tier compared with Gemini 2.0 Flash.
- Increases maximum output length from 8,192 to 65,536 tokens compared with Gemini 2.0 Flash.
- Published evaluations report stronger reasoning and multimodal performance than Gemini 2.0 Flash.
Known flaws & limitations
- Can generate factual errors and unsupported claims despite reasoning or grounding.
- The base gemini-2.5-flash model generates text, not native image or audio output.
- Its documented knowledge cutoff is January 2025; later information requires external sources.
- Thinking tokens count toward output usage and can increase latency and charges.
- The listed standard paid-tier input price is USD per million text, image or video tokens; audio input has separate pricing.
Risks, safety & security issues
- Untrusted documents and tool responses can contain prompt-injection instructions; authorization must be enforced outside the model.
- Incorrect or biased responses can cause harm in consequential decisions without qualified review.
- Sensitive information included in prompts or exposed through tools can create privacy and confidentiality risks.
- Malicious packages and fraudulent AI-credit offers can compromise developer systems independently of model safeguards.
Good fit for
- High-volume document summarization, extraction and classification.
- Multimodal analysis of documents, images, audio and video.
- Tool-enabled assistants with narrowly scoped permissions and validated outputs.
- Workflows that benefit from adjustable reasoning effort and long-context input.
Poor fit for
- Autonomous medical, legal or financial decisions without qualified review.
- Native image or audio generation using this base model.
- Unrestricted execution of actions derived from untrusted documents or tool responses.
- Current-events answers without external evidence.
References
Profile last reviewed 10/24/2026 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.