Google launches Gemini 3.8 Flash at 3.7 pricing, adds restricted Cyber variant
- Google released Gemini 3.8 Flash at the same introductory API price as 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens.
- Google says Gemini 3.8 Flash Cyber has frontier-level vulnerability detection and automated patching performance, but limits access to trusted defenders in its Fairwind Program.
- Both releases use the same foundational model and long-running agentic loops that recursively evaluate and refine the models; Google says cybersecurity training contributed to coding and reasoning gains.
- Google says 3.8 Flash beats most larger frontier models on DeepSWE v1.1 for end-to-end software engineering and scores 54.9% on HLE-Verified.
- The release is Google's third Flash model update in six weeks, following Gemini 3.7 Flash three weeks earlier.
Hacker News opinions
The post is 404 now, though I can still use the model in the Gemini chat website.
I found archived copies of both the page and the PDF model card. Google's original post is gone, so use those if you want to inspect the claims.
It feels like Google has stopped shipping frontier models for external consumption and is concentrating on selling compute.
I don't see how that follows when a Flash model is reportedly close to Opus and Sol on some benchmarks.
My guess is Gemini 3.5 Pro gets skipped and Gemini 4 Pro is next. Gemini 4 pre-training is reportedly underway, although I don't know anything beyond public information.
A month is too short for a major organization-wide turnaround. These releases likely come from workstreams that began under the previous DeepMind leadership.
The model card's safety chart is sloppy: a comparison column labeled 3.8 versus 3.7 uses Gemini 3 Flash for tone and instruction-following. That makes 20% of the comparison misleading.
Flash 3.7 was extremely fast in my tests, but Sol did much better on production code in any non-tiny codebase. I use Flash for prototypes rather than serious codebase work.
The model is very selective about benchmarks. TerminalBench is only 10%, and I have seen it overreact to trivial prompts by building an entire multi-panel app.
DeepSWE currently puts it above Opus 5, and Artificial Analysis gives it an intelligence score of 59. For a Flash model, that is impressive, but I want to see real use before trusting the numbers.
A 59 score is not equal to Opus 5's 63 at maximum effort, and Opus can use about four times fewer output tokens. The speed comparison is less clear once token use is included.
DeepSWE is public, so contamination is a concern. It is also far behind Opus 5 on TerminalBench 4.0, 19.1% versus 51.8%.
I tried it on token-heavy coding work. It handles many tasks, but it needs more prompting and supervision than Claude or Sol, so it can be slower in practice without an auto-approve mode.