Google's Gemini 3.8 Flash model card reports a DeepSWE v1.1 score of 73.7%, up from 65.3% for Gemini 3.7 Flash. The model builds on 3.7 Flash and retains adjustable reasoning effort, allowing developers to trade response quality against latency and cost.
The card, dated September 2, lists text, image, audio and video inputs within a context window of up to one million tokens. Output is text, with a 64,000-token limit. Google lists distribution through its Gemini API, AI Studio, Gemini app and other products.
The safety results are not uniformly better. Google's automated multilingual safety evaluation regressed by 5.4 percentage points against 3.7 Flash, on a measure where lower is better. Google says manual review found most flagged losses were false positives or non-egregious cases.
The card also records hallucinations, occasional slow responses or timeouts, and higher token use at greater effort levels. Its coding gains are Google's reported evaluation results, not evidence that every agent workflow becomes faster or more reliable.