Google released Gemini 3.8 Flash on September 2, its third Flash model in six weeks. The timing is worth noting: it lands on the same day as Anthropic’s Fable 5.1. Two vendors, one day, both claiming a jump in coding.
The numbers Google cites aim straight at that target. On Terminal-Bench 2.1, 3.8 Flash hits 90.8 percent, up from 81.6 percent for its predecessor 3.7 Flash. On DeepSWE v1.1, a benchmark for long-horizon software work, Google says the model beats most larger frontier models at solving complete engineering problems end to end, and at a fraction of the cost. Google calls it its most intelligent workhorse and points to gains in multi-step reasoning, multi-file refactoring, and deterministic tool execution.
The price stays where it was: $0.75 per million input tokens, $3.75 for output. That’s the actual point of the Flash line. Not the absolute top score, but performance close to the big models at a price where you can run the thing on high-volume workloads. If you’re building agents that fire thousands of times, you count cost per task, and there a cheap model with solid coding quality beats an expensive flagship.
There’s also a second variant: Gemini 3.8 Flash Cyber. It replaces the older 3.5 Cyber release and is available only to vetted security researchers, government agencies, and critical infrastructure operators, through Google’s new Fairwind Program. The pattern is familiar by now: offensive cyber capabilities get locked behind approval. OpenAI just did this with Astra, and Anthropic reserves its sensitive exploit capabilities for its more restricted models. Google is following suit with a program of its own.
I’d read the benchmark numbers with the usual caution. They come from Google, and a 90.8 percent on a single benchmark says little about how the model feels in daily use. The more telling comparison comes over the next few days, once independent people put Fable 5.1 and Gemini 3.8 Flash side by side on real tasks. Simon Willison already updated his llm-gemini plugin to 3.8 Flash the same day, complete with three thinking levels: low, medium, and high. The race in the cheap coding tier is wide open in a way it hasn’t been for a while.
Sources: 9to5Google: Gemini 3.8 Flash launch · DataCamp: Gemini 3.8 Flash · Simon Willison: Release llm-gemini 0.34