Google DeepMind's Gemini 3.8 Flash Beats Claude Opus 5 at 6x Lower Cost
Google's third Flash release in as many months pushes a workhorse model into frontier-tier territory on agentic coding, legal, and finance benchmarks.
- Google released Gemini 3.8 Flash, the third Flash iteration in about three months.
- Live today for Gemini app Pro and Ultra users, plus AI Studio, Antigravity, and the Gemini API.
- Priced at $0.75 input and $3.75 output per million tokens through end of 2026.
- Beats Opus 5 and GPT-5.6 Sol on Vals Finance Agent v2, Harvey Legal Agent Bench, and Terminal-bench 2.1.
- Big jumps over 3.7 Flash on Terminal-bench 4.0 (19.1% vs 11.2%) and OSWorld-2.0 (59.0% vs 50.6%).
- Slight regression in non-English safety per Google's model card.
Google DeepMind has released Gemini 3.8 Flash, the third refresh of its mid-tier Flash line in roughly as many months. It's live today for Gemini app subscribers on Pro and Ultra plans, and available inside Google AI Studio, the Gemini API, and Google Antigravity for developers.
Gemini 3.6 Flash landed in July, 3.7 Flash followed in mid-August, and 3.8 Flash has arrived just weeks later. Each release has pushed benchmark scores without touching the pricing. The goal with 3.8 Flash is the same as before: keep latency and cost low while closing the gap with frontier-tier reasoning. The recurring theme in the release notes is agentic reliability: fewer dropped tool calls, tighter multi-step reasoning, and cleaner output.
What actually changed
Google positions 3.8 Flash as its strongest Flash model yet for mid-difficulty long-horizon software engineering, autonomous agents, and enterprise workflows. The biggest reported gains over 3.7 Flash:
- Terminal-bench 4.0 jumps to 19.1% from 11.2%, a general agentic benchmark where cheaper models typically collapse.
- BioMysteryBench Human Difficult climbs to 56.5% from 43.5%, and OSWorld-2.0 rises to 59.0% from 50.6%.
- HLE-Verified, a multidisciplinary expert reasoning eval, reaches 54.9%.
Pricing that undercuts the flagship tier
At $0.75 per million input tokens and $3.75 per million output tokens, the introductory rate runs through the end of 2026, after which pricing moves to $1.50 input and $7.50 output. That puts it roughly 6-7x cheaper than Claude Opus 5 ($5 input / $25 output) and well below GPT-5.6 Sol ($4 input / $20 output). On several benchmarks, 3.8 Flash beats both.
Where it wins, where it doesn't
The strongest results are in domain-heavy agent work. On the Vals Finance Agent V2 eval, 3.8 Flash posts a 10.0% pass rate, more than double GPT-5.6 Sol's 2.5% and ahead of Opus 5's 6.7%. It also outperforms both on Harvey's Legal Agent Benchmark.
Coding follows a similar pattern. On DeepSWE v1.1, a long-horizon software engineering benchmark, 3.8 Flash outperforms most larger frontier models at a fraction of the cost. It also edges out Opus 5 on Terminal-bench 2.1 with 89.4% against Opus 5's 89.1%. Frontier reasoning models still hold an advantage on the very hardest general-agent benchmarks, and Google's own model card flags one regression worth noting: safety performance across non-English languages slipped slightly relative to 3.7 Flash.
Who should switch
3.8 Flash is the obvious candidate if you run agents at volume, build long-horizon coding loops, or process documents at scale where Opus-class pricing breaks the unit economics. On the Gemini API, swapping the model string is enough to A/B test it against your current stack. If your pipeline depends on non-English safety behavior, run your own regression tests before flipping to production.
Comments
