Gemini 3.8 Flash

Gemini 3.8 Flash landed on 2 September 2026, three weeks after 3.7 Flash. Google kept the pitch the same: speed and price rather than peak intelligence, with the gains concentrated in software engineering, agentic tasks, and multi-step reasoning.

The price to watch

$0.75 per million input tokens and $3.75 per million output, which is what 3.7 Flash cost. The number to put in your calendar is 1 January 2027, when the introductory rate ends and both figures double to $1.50 and $7.50. If you’re sizing a workload on Flash economics, size it on the 2027 price.

The Cyber variant

Google shipped a sibling, 3.8 Flash Cyber, on the same day. It runs with deliberately looser mitigations for security work and is restricted to trusted defenders through the new Fairwind Program, which covers government authorities, critical infrastructure operators, and software maintainers who apply and get approved. It isn’t something a reader can pick in the Gemini app, which is why it doesn’t get its own entry here.

Where it fits

On raw capability, Google has been left behind by the frontier models from OpenAI and Anthropic. That matters less than the benchmark tables imply. How clever the model is turns out to be only one part of a working setup. The other part is the software wrapped around it: what it can see, what it can reach, which of your other tools it can act on. In AI circles that wrapper gets called the harness, and for day-to-day work it’s often worth more than the last few points of raw capability.

That’s why Flash works as a coworker for me specifically. I use a lot of Google services, so the model already sits next to the things I work on, and inside Antigravity it’s inexpensive, fast, and capable enough. It’s also become excellent at watching and reviewing video, which is a job most models are still bad at.

The rest of the advice for Flash hasn’t changed: use it for volume, and know where the volume ends. For hard logic or high-stakes reasoning, step up to a frontier model and let Flash handle everything around it.

The Verdict

Best for: High-volume execution work: coding, agents, copy, and anything where speed and price beat peak intelligence.