Google releases Gemini 3.8 Flash, its third Flash model in six weeks

Ryan Whitwam notes Google has not shipped a frontier Gemini Pro since early 2026, but is on its third Flash in six weeks, making Gemini 3.5 Pro look less likely. Google calls 3.8 Flash its best reasoning and coding model. Two variants: a 'workhorse' Flash for agents and software work, and Flash Cyber tuned for vulnerability detection and mitigation.

API access is at an introductory rate through year-end: $0.75 / $3.75 per million input/output tokens, then $1.50 / $7.50 — though Whitwam figures new models will arrive before the hike. He reads the discount as necessary because other labs have cut token prices. Benchmarks show marginal gains over 3.7 in most tests, larger in coding; 3.8 tops DeepSWE at the discounted rate. Google reportedly delayed 3.5 Pro when coding lagged; if these numbers hold, even Flash is competing with market leaders. On OSWorld-2.0 computer use, 3.8 improves on 3.7 but remains far behind Claude Opus; GPT is also weak there. Cyber replaces the 3.5 Cyber: Chrome security saw 2.6x patch accuracy; Cloud found a critical vulnerability in two hours; Wiz and Palo Alto Networks supplied statements. Flash is across Google's ecosystem today for Pro/Ultra in the Gemini app, with AI Studio free to tinker. Cyber stays with trusted testers and governments.

Whitwam's frame is product-line abandonment. Pro is paused; Flash is the train. That is not necessarily failure — it can be a decision that the money is in a cheap, fast coder that updates every two weeks. It does mean 'Gemini 3.5 Pro' as a promised frontier object is being talked out of existence by the shipping schedule.

The OSWorld paragraph is the one Google would not lead with. Coding boards are where Flash looks like a giant-killer. Computer use is where it still is not Claude. Anyone buying 'autonomous agents' off this launch should read that gap as the actual product limitation.

Signal Desk's read: introductory pricing through December, with a doubled list price on the calendar, is a customer-acquisition coupon on a model family that refreshes faster than the coupon. Whitwam is right that the hike may never land on this SKU. The likelier path is 3.9 at another 'intro' rate. Cyber's partner quotes (Wiz, Palo Alto) are the enterprise-security distribution strategy: do not GA a hacking model, seed it with the firms that already sell defense.

If 3.5 Pro was delayed for coding, and Flash now leads DeepSWE, Google may never need the bigger SKU for the audience that writes pull requests. That would be a real strategy, and it would leave multimodal-frontier buyers waiting.

Google's Flash line has absorbed the release energy that used to go to Pro. Anthropic and OpenAI are also racing on price. Computer-use evals remain a separate, still-unsolved product from coding evals, and Ars is one of the few writeups that keeps them apart.