Gemini 3 Flash Preview
google-deepmind/gemini-3-flash-preview
Creator
Google DeepMindReleased
2025-12-17Weights
not reportedContext
not reportedPrice in / out per Mtok
not reported / not reportedState
rankedHealth summary
10 of 15 scores sit on benchmarks that still discriminate. Every number here has been measured independently. 9 sit on benchmarks with a high risk of test-set contamination.
Scores
| Benchmark | Config | Score | Harness | Provenance | Measured | Health |
|---|---|---|---|---|---|---|
| Knowledge | ||||||
| GPQAnearing saturation | base | 89.4% | not reported | ind. | 2025-12-17 | supersededdisputed |
| Coding | ||||||
| Terminal-Benchactive | base | 64.3% | Junie CLI | ind. | 2025-12-17 | contaminated |
| Terminal-Benchactive | base | 51.7% | Terminus 2 | ind. | 2025-12-17 | contaminated |
| Terminal-Benchactive | base | 51.0% | Gemini CLI | ind. | 2025-12-17 | contaminateddisputed |
| Reasoning | ||||||
| ARC-AGI-2nearing saturation | base | 33.6% | not reported | ind. | 2025-12-17 | supersededdisputed |
| On benchmarks that no longer discriminateKept visible on purpose. These are often the numbers a launch announcement leads with. | ||||||
| SWE-bench Verifiedsaturated | base | 75.4% | not reported | ind. | 2025-12-17 | saturatedcontaminated |
| ARC-AGIsaturated | base | 84.7% | not reported | ind. | 2025-12-17 | saturatedcontaminatedsupersededdisputed |