Google highlights four builder projects made with Gemini 3.8 Flash

Google showcased four projects made with Gemini 3.8 Flash on September 28, 2026, including satellite-path visualisation, animated art and interactive models. The examples give developers a view of how creators are using the model, but the published descriptions do not establish whether the outputs are accurate or reliably reproducible across other tasks.

The showcase follows Google's September 2 introduction of Gemini 3.8 Flash. At launch, the company said the model improved on Gemini 3.7 Flash in software engineering, agent tasks and multistep reasoning. Those are Google's performance claims; the four featured projects illustrate uses of the model rather than provide controlled tests of those claims.

Satellite paths and animated Seigaiha waves

Another featured creator, Noctus, used the model to turn Seigaiha waves, a traditional Japanese pattern, into an animation resembling a moving ink painting. Noctus described the result as convincing. The example concerns a visual transformation: neither the creator's description nor Google's showcase supplies a broader measure of the model's performance on animation tasks.

A T. rex skeleton and a transmission prototype

Google says Emily, who posts as @IamEmily2050, asked Gemini 3.8 Flash to make a T. rex skeleton through a four-phase prompt with strict requirements, banned shortcuts and accuracy checks. The creator's posted prompt describes a procedural Three.js skeleton for the specimen Sue. The prompt documents the challenge, but its text does not independently establish the anatomical accuracy of the resulting model.

Hakm challenged the model to build an interactive automatic-transmission simulation. Google describes a prototype with 10 camera views, four display modes, labels and a side panel that explains a part when clicked. The creator's post separately confirms the challenge, while Google's account is the source for the full list of prototype features. The available posts do not constitute a technical audit of the simulation.

What the performance figures show

Google's model card reports a 73.7% score for Gemini 3.8 Flash on DeepSWE v1.1, compared with 65.3% for Gemini 3.7 Flash and 74.0% for Claude Opus 5 on the card's listed evaluation. The same card reports 19.1% on Terminal-bench 4.0 for Gemini 3.8 Flash, against 51.8% for Claude Opus 5 and 37.3% for GPT-5.6 Sol. These are vendor-reported results on different tests, not scores for the four showcased builds.

A separate GenBench report from September 16 recorded an 87.6 average for Gemini 3.8 Flash across three lanes without media input, using OpenRouter. It recorded $2.30 in cost and nine minutes of parallel wall-clock time for that run. GenBench listed comparison scores of 90.7 for Grok 4.5 high, 87.9 for GPT-5.6 Terra high and 85.9 for Kimi K3 within its setup.

GenBench said its run excluded image, video, audio and voice providers, and described the outcome as a routing signal rather than certification for production use. Its report said all three lanes missed or lacked strict gates. The result therefore offers a limited point of comparison for developers; it cannot validate the multimodal art project or the accuracy of the skeleton and transmission models.

What developers can take from the showcase

The examples identify concrete tasks creators have tried: plotting orbital paths, animating a pattern and constructing interactive 3D models. Google's published account and the creators' posts support those descriptions to different degrees. They leave open how often similar prompts would succeed, whether the technical details of the outputs hold up to expert scrutiny and how much work each creator did beyond the prompts described.

Google lists hallucinations, occasional slowness or timeouts, and higher token use at greater effort levels among the model's limitations. Those cautions are especially pertinent when an output appears to depict real positions or technical anatomy and machinery. The four projects show possible workflows, while the available reporting does not establish that their underlying data or engineering details have been independently verified.