Google Gemini 3.8 Flash Has a Mix of Improved Benchmarks

admin
By
3 Min Read

Gemini 3.8 Flash is Google’s newest Flash-tier model released September 2, 2026. It is the most capable Flash model yet for mid-difficulty long-horizon software engineering, autonomous agents, and enterprise workflows, while keeping Flash-level speed and the same introductory pricing as 3.7 Flash. It is a value model.

Pricing is the same as 3.7 Flash

Introductory (through Dec 31, 2026): $0.75 / 1M input, $3.75 / 1M output
Regular (from Jan 1, 2027): $1.50 / $7.50

Where it is strong
Vals Finance Agent v2 — 61.4% (best)
Harvey’s Legal Agent Benchmark — 10.0% (best. all models score poorly here)
Terminal-bench 2.1 — 89.4% (narrowly best)
CharXiv Reasoning (charts, no tools) — 86.2% (best)
LVBench long video — 87.8% agentic / 87.1% static (best)
HLE-Verified — 54.9% (best)
BioMysteryBench (Human Difficult) — 56.5% (best)
LABBench2 — 86.2% (best)

These wins cluster in multimodal understanding (video + charts), domain-specific professional agents (finance, legal, biology), and mid-difficulty terminal/agentic coding.

Where it is behind

DeepSWE v1 (hard long-horizon SWE) — 71.0% vs Opus 5’s 74.0% and GPT-5.6 Sol’s 72.7%
GDPVal-AA v2 knowledge-work Elo — 1545 vs Opus 5’s 1824
Terminal-bench 4.0 (harder general agents) — 19.1% vs Opus 5’s 51.8%
OSWorld-2.0 computer use — 59.0% vs Opus 5’s 75.4%
GDP.PDF — 35.0% vs GPT-5.6 Sol’s 40.0%

Share This Article
Leave a Comment

Leave a Reply