Gemini 3.8 Flash beats flagships on some benchmarks, not all

Google put Gemini 3.8 Flash into Agent Studio on GCP on September 2, and this is now the third Flash release in roughly five weeks — 3.6 landed July 21, 3.7 followed August 13, and now this one, a release cadence closer to a team shipping hotfixes than a company rolling out flagship models. The context window sits at just over 1 million tokens, and Google built this version specifically around agentic workflows: multi-step tasks where the model breaks a request into a sequence of actions and runs through them without a person checking in after every step.

The benchmark story is genuinely good for Google, just not in the blanket way it’s being described online. On Vals Finance Agent v2, a benchmark built around financial-analyst tasks, Gemini 3.8 Flash scored 61.4% against GPT-5.6 Sol’s 53.8% and even beat Opus 5’s 58.6%. On Harvey’s Legal Agent Benchmark, covering complex legal workflows, it hit 10.0% versus GPT-5.6 Sol’s 2.5%. Those are real, specific wins on tasks that map to how people actually use these models day to day, not synthetic reasoning puzzles nobody encounters in practice.

It’s not sweeping the board, though. Terminal-bench 4.0, a harder general-agent test, has Gemini 3.8 Flash at 19.1% against Opus 5’s 51.8% — not close. OSWorld-2.0, which tests computer-use tasks, shows a similar gap: 59.0% versus 75.4%. Fable 5 still holds a higher ceiling on the hardest reasoning tasks generally, and GPT-5.6 Sol remains stronger on general coding outside the specific agent benchmarks where Gemini 3.8 Flash wins.

What’s actually happening is narrower and arguably more interesting than “Flash beats everyone”: a cheap, fast model is now competitive with, and sometimes ahead of, full-size flagships on the specific agentic tasks it was built for, while still losing clearly on the tasks it wasn’t optimized around. Google has now shipped that improvement three times in five weeks, a pace that makes today’s benchmark table a fairly short-lived snapshot. DeepSeek made a similar case for itself with V4-Pro not long ago, closing the gap with flagship models on specific benchmarks while still trailing on others, and the pattern across the industry right now looks less like one model pulling ahead of everyone and more like every lab finding its own narrow lane to win in.