gemini-3-flash
Meant to excel at
No sourced claims yet.
Watch out for
No sourced claims yet.
Benchmarks
Agentic coding percentile 0.14
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| swe-bench pro | 49.6 | 0.00 | Vendor's own site deepmind.google | 2026-05-19 |
| swe-bench pro (scale seal public set) | 34.6 | 0.27 | Independent evaluator morphllm.com | 2026-09-14 |
| swe-bench verified | 78 | 0.20 | Vendor's own site blog.google | 2025-12-17 |
| terminal-bench 2.1 | 58 | 0.09 | Vendor's own site deepmind.google | 2026-05-19 |
Computer use percentile 0.25
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| osworld-verified | 65.1 | 0.25 | Vendor's own site deepmind.google | 2026-05-19 |
Reasoning percentile 0.04
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| gpqa diamond | 90.4 | 0.00 | Vendor's own site blog.google | 2025-12-17 |
| humanity's last exam | 33.7 | 0.08 | Vendor's own site blog.google | 2025-12-17 |
Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.
Practitioner reports
- +/-
A reviewer found Gemini 3 Flash produced a working snake game from one prompt (where Gemini 3 Pro's first try failed) and fixed Pro's broken version, but rated it good only for simple coding and advised Pro for complex tasks. NOTE: review last updated 2026-01-13, older than the 120-day window.
Task: feature | Compared with gemini-3.1-pro | 2026-01-13 older than 90 days
Practitioner report cybernews.com