Gemini 3.8 Flash scorecard
Section headers group the seventeen rows by capability. The page intentionally compares Gemini 3.8 Flash against itself only; every other model on the site compares against Gemini 3.8 Flash instead.
| Benchmark | Gemini 3.8 Flash |
|---|---|
| Agentic | |
| Terminal-Bench 2.1 | 89.4% |
| Terminal-Bench 4.0 | 19.1% |
| OSWorld 2.0 | 59.0% |
| Vals Finance Agent v2 | 61.4% |
| Harvey Legal Agent | 10.0% |
| LVBench (agentic) | 87.8% |
| Coding | |
| DeepSWE v1.1 | 73.7% |
| SWE-Bench Pro | 61.6% |
| GDP.PDF | 35.0% |
| Reasoning | |
| HLE-Verified | 54.9% |
| LVBench (static) | 87.1% |
| CharXiv Reasoning | 86.2% |
| LAB-Bench 2 | 86.2% |
| BioMysteryBench (difficult) | 56.5% |
| AA-LCR (long context) | 82.0% |
| Knowledge | |
| GDPval-AA v2 | 1545 Elo |
| AA Intelligence Index | 59 |
How to read the table
- Source spread: Terminal-Bench 2.1 is reported at 89.4% by four sources and 90.8% by DataCamp; DeepSWE v1.1 is reported at 73.7%, 71.0% and 73.8%. The 89.4% figure is the majority; the 90.8% is the outlier.
- GPQA Diamond at 95.3% is reported only by BenchmarkList. Google's own announcement does not cite it, so it is not corroborated.
- CWE-Bench pass@1 at 47.2% (Cyber variant) is measured against a frontier model at 47.8%. Quoting 47.2% alone overstates it.
- Reasoning levels: AA Intelligence Index is 59 at high, 57 at medium and 52 at low. Cost per AA task falls from $0.58 to $0.41 to $0.24 across the same three levels.
- Effort tiers and harness versions differ across evaluators. Treat each row as its own evaluator's measurement, not a like-for-like across the table.
Ready to test the workflow?
Why 3.8 Flash scores low on some rows
Terminal-Bench 4.0 measures open-ended long-horizon agent work where the model chooses its own next step; Gemini 3.8 Flash scores 19.1% there. OSWorld 2.0 measures computer use; 3.8 Flash scores 59.0%. Computer use is listed as a capability but is not this model's strength — Google itself recommends staying on 3.7 Flash for efficiency-first workloads.