Gemini 3.8 Flash

1. Coding is where 3.8 Flash really starts to stand out

This is probably the biggest reason to pay attention to the new model.

On DeepSWE v1.1, Gemini 3.8 Flash scored 73.7%, compared with 65.3% for 3.7 Flash and 74.0% for Claude Opus 5.

On Terminal-Bench 2.1, it scored 89.4%, slightly ahead of Opus 5 at 89.1%.

That puts Flash in a very different conversation from the old “fast model for simple tasks” idea. Complex coding, terminal work, and agent-style development are clearly part of the target now.

2. It’s getting better at specialized work

Finance and legal tasks are two good examples.

On Vals Finance Agent V2 and Harvey's Legal Agent Benchmark, 3.8 Flash scored higher than both 3.7 Flash and Claude Opus 5.

It also scored 54.9% on HLE-Verified, which covers questions across STEM, humanities, and professional fields.

That doesn't mean it can suddenly solve every expert-level problem. The more useful takeaway is that it has become better suited to tasks where the answer takes several rounds of reasoning to get right.

3. You can control how hard it thinks

Gemini 3.8 Flash gives developers three reasoning levels: low, medium, and high thinking effort.

In a third-party intelligence test, the model scored 59 at high effort, 57 at medium, and 52 at low.

That gives developers a useful trade-off. Simple requests don't need maximum reasoning. For complicated code or longer agent workflows, you can turn it up and spend more tokens when the extra work is actually worth it.

4. It thinks more without completely giving up its speed

3.8 Flash can generate around 300 tokens per second, roughly 4–6 times faster than many mainstream models.

Its p50 latency is also about 52% faster than 3.6 Flash.

So Google hasn't simply traded speed for reasoning. Flash is still fast. It just has more room to think when the task calls for it.

⚡ How It Has Changed From Gemini 3.7 Flash

This isn't a ground-up rebuild of 3.7 Flash.

Google is still building on the same general technical foundation, including the model architecture, training data, hardware, and software stack.

The bigger change is how the model handles a difficult job.

Instead of rushing toward an answer, 3.8 Flash can spend more time reasoning, call tools, check the results, and keep iterating. Across 14 benchmark groups, it took first place in 8.

There is a price for that, though: a single task costs about 40% more than with 3.7 Flash.

And that makes the upgrade a lot easier to judge.

For a quick function or simple question, you may not notice much. For a large codebase, a long-running agent, or a complicated analysis task, those extra reasoning steps can be much more useful.

The good stuff

• Stronger performance in software engineering, agents, and professional tasks
• 73.7% on DeepSWE v1.1 and 89.4% on Terminal-Bench 2.1
• Around 300 tokens/sec output speed
• 1M-token context window, enough for roughly 1,500 pages of A4 text
• Three thinking-effort levels
• API promotional pricing matches 3.7 Flash through the end of 2026

The trade-offs

The biggest thing that gives me pause is the price.

Harder tasks already tend to use more tokens, and 3.8 Flash costs about 40% more per task than 3.7 Flash. If you're running agents all day, that difference can show up pretty quickly on the bill.

There's also the access issue. In the Gemini app, 3.8 Flash is limited to Pro and Ultra. API promotional pricing ends on December 31, 2026, after which both input and output prices double.

And then there's the gap between benchmarks and real work.

Benchmarks can look great on paper. Real projects are another story.

One evaluation found that 3.8 Flash could generate runnable 3D scene code in two or three minutes, but the final result could still be a case of “it runs, but it’s not right.”

That's a useful reminder: a high benchmark score doesn't guarantee a clean result when you drop the model into a real project.

Who should use it?

If you spend a lot of time on complex coding, AI agents, or tasks that require several rounds of reasoning, 3.8 Flash is worth a look.

It's also interesting for finance, legal, and other professional workflows where the model needs to work through information step by step.

For businesses, the math is pretty simple: if the extra reasoning saves enough human time, the higher token cost may be worth it.

Who probably doesn't need it?

If you're mainly using AI for chat, summaries, or basic rewriting, 3.8 Flash is probably more power than you need.

Developers watching API costs closely don't need to rush into an upgrade either. A lot of straightforward tasks can still be handled perfectly well by 3.7 Flash.

And if you just want to try it without paying for Pro, waiting is a perfectly reasonable option.

Google's direction here is pretty clear to me: Flash is no longer trying to be just the fast and cheap model.

It's starting to go after the harder jobs, too.

That's a good thing for users. Just keep one thing in mind: the better the model gets at thinking, the smarter your bill may get at finding your wallet.