Google DeepMind's generally available Flash model for long-horizon software engineering, autonomous agents, and complex enterprise workflows.
Terminal-Bench 4.0 in the Artificial Analysis evaluation harness.
Humanity's Last Exam as evaluated by Artificial Analysis; the source configuration is preserved.
GDPval-AA v2.1 professional knowledge-work Elo. This is a task lens, not an assessment of every research skill.
GDP.pdf professional document understanding, including visual tables, charts and diagrams; not an audio/video score.
AutomationBench-AA multi-step business workflows in the Artificial Analysis harness.
Source-backed rows with the original evaluation context preserved.
| Benchmark | Score | Source | Evaluation notes |
|---|---|---|---|
| AutomationBench-AAChecked 2026-10-02 | 60% | Artificial Analysis | Artificial Analysis Intelligence Index v4.3.2; configuration=high. Rounded public score; checked date is not the evaluation execution date. |
| AA GDP.pdfChecked 2026-10-02 | 21% | Artificial Analysis | Artificial Analysis Intelligence Index v4.3.2; configuration=high. Rounded public score; checked date is not the evaluation execution date. |
| GDPval-AA v2.1Checked 2026-10-02 | 1412 | Artificial Analysis | Artificial Analysis Intelligence Index v4.3.2; configuration=high. Rounded public score; checked date is not the evaluation execution date. |
| AA Humanity's Last ExamChecked 2026-10-02 | 48% | Artificial Analysis | Artificial Analysis Intelligence Index v4.3.2; configuration=high. Rounded public score; checked date is not the evaluation execution date. |
| Artificial Analysis Intelligence Index v4.3.2Checked 2026-10-02 | 41 | Artificial Analysis | Artificial Analysis Intelligence Index v4.3.2; configuration=high. Rounded public score; checked date is not the evaluation execution date. |
| AA Terminal-Bench 4.0Checked 2026-10-02 | 20% | Artificial Analysis | Artificial Analysis Intelligence Index v4.3.2; configuration=high. Rounded public score; checked date is not the evaluation execution date. |
| Artificial Analysis Intelligence IndexEvaluated 2026-09-03 | 59 | Artificial Analysis | Artificial Analysis Intelligence Index v4.1.1; Gemini 3.8 Flash high reasoning configuration. |
| Artificial Analysis Intelligence IndexEvaluated 2026-09-03 | 57 | Artificial Analysis | Artificial Analysis Intelligence Index v4.1.1; Gemini 3.8 Flash medium reasoning configuration. |
| Artificial Analysis Intelligence IndexEvaluated 2026-09-03 | 52 | Artificial Analysis | Artificial Analysis Intelligence Index v4.1.1; Gemini 3.8 Flash low reasoning configuration. |
| BioMysteryBench, human-difficultEvaluated 2026-09-02 | 56.5% | Google DeepMind | Google-reported BioMysteryBench result; task set=human-difficult. |
| BioMysteryBench, human-solvableEvaluated 2026-09-02 | 88.8% | Google DeepMind | Google-reported BioMysteryBench result; task set=human-solvable. |
| CharXiv RQEvaluated 2026-09-02 | 86.2% | Google DeepMind | Google-reported CharXiv RQ result; evaluation settings are source-specific. |
| DeepSWE v1.1Evaluated 2026-09-02 | 73.7% | Google DeepMind | Google-reported DeepSWE v1.1 result; highest listed thinking level=high; preserve the mini-SWE-agent harness context. |
| Vals Finance Agent v2Evaluated 2026-09-02 | 61.4% | Google DeepMind | Google-reported Vals Finance Agent v2 result; tools enabled. |
| GDP.pdfEvaluated 2026-09-02 | 35% | Google DeepMind | Google-reported GDP.pdf result. |
| GDPval-AA v2Evaluated 2026-09-02 | 1545 | Google DeepMind | Google comparison-table GDPval-AA v2 score; score unit is Elo. |
| GPQA / GPQA DiamondEvaluated 2026-09-02 | 95.3% | Google DeepMind | Google comparison-table GPQA Diamond result. |
| Harvey Legal Agent BenchmarkEvaluated 2026-09-02 | 10% | Google DeepMind | Google-reported Harvey Legal Agent Benchmark result; score is complete-task resolution, not per-criterion accuracy. |
| HLE-VerifiedEvaluated 2026-09-02 | 54.9% | Google DeepMind | Google-reported HLE-Verified result on the verified set. |
| LABBench2Evaluated 2026-09-02 | 86.2% | Google DeepMind | Google-reported LABBench2 result. |
| LVBenchEvaluated 2026-09-02 | 87.8% | Google DeepMind | Google-reported LVBench result with tools; context and video sampling settings are source-specific. |
| OSWorld 2.0Evaluated 2026-09-02 | 59% | Google DeepMind | Google-reported OSWorld 2.0 result; task release and computer-use harness are source-specific. |
| SWE-bench ProEvaluated 2026-09-02 | 61.6% | Google DeepMind | Google comparison-table result; benchmark and harness details remain source-specific. |
| Terminal-Bench 2.1Evaluated 2026-09-02 | 89.4% | Google DeepMind | Google-reported Terminal-Bench 2.1 result; preserve the Terminus 2 harness context. |
| Terminal-Bench 4.0Evaluated 2026-09-02 | 19.1% | Google DeepMind | Google-reported Terminal-Bench 4.0 result; do not compare directly with Terminal-Bench 2.1. |
Official model card for Gemini 3.8 Flash model lineage, 1M-token context, 64K output limit, March 2026 knowledge cutoff, safety notes, and links to the evaluation methodology.
Official API documentation for the gemini-3.8-flash model ID, GA status, reasoning levels, limits, and introductory pricing.
Underlying raw metrics and exact model configurations remain authoritative.
Underlying raw metrics and exact model configurations remain authoritative.