Google Showcases Four Developer Builds Made With Gemini 3.8 Flash

Google used its blog to highlight four projects built with Gemini 3.8 Flash, the model it released earlier in September and describes as its most intelligent workhorse. The post frames the model’s gains as coming from a design choice rather than raw scale: on complex tasks, Gemini 3.8 Flash runs extra reasoning steps and calls tools iteratively before settling on an answer, which Google says brings it closer to the performance of costlier frontier models while keeping Flash-tier latency.

The four examples span quite different kinds of work. One developer used the model with Google Antigravity to map the live paths of satellites, space stations and orbital rockets. Another used its multimodal handling to turn a traditional Japanese wave pattern into an animation resembling a moving ink painting. A third asked the model to build an anatomically accurate dinosaur skeleton through a four phase prompt with banned shortcuts and accuracy checks, testing how far the model would push on multistep STEM reasoning. A fourth built an interactive automatic transmission simulator with ten camera views, four display modes and a side panel that explains each part on click, an engineering task that depends on the model tracking many moving pieces at once.

None of the four builds is presented as an official product; they are community demonstrations Google chose to spotlight to show what the model can be asked to do beyond a chat window.

The numbers

Google’s own API price table lists Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 per million output tokens under a promotional rate that runs through the end of 2026, after which both figures roughly double. The model supports a context window of 1,048,576 input tokens, matching the million token window Google has standardised across its current Gemini lineup, alongside a maximum output of 65,536 tokens.

Those numbers matter next to the four projects Google chose to show off. A satellite tracker or a transmission simulator that calls the model repeatedly to reason through steps, rather than answering in one pass, spends more tokens per task than a single query would suggest. At $0.75 per million input tokens, a workhorse model that iterates several times before finishing a task can still cost a fraction of Gemini 3.1 Pro’s $2.00 rate, which is the choice this roundup is really arguing for: pay less per token and let the model work harder, rather than pay more for a single-pass answer from a pricier model.