Gemini 3.8 Flash vs 3.7 Flash : Nouveautés et faut-il mettre à jour ?

Gemini 3.8 Flash, launched by Google on September 2, 2026, offers several enhancements over its predecessor, 3.7 Flash. The main difference lies in the model's operation, which involves shorter step reasoning, self-checking of results, and iterative tool calls. While the scores improve, each task may consume more tokens. The main question is whether the quality gain justifies the additional cost for a workload and whether the code supports the API changes.

Google has stated that Gemini 3.7 Flash remains fully supported with no announced deprecation date. Migration is not urgent. The pricing and constraints for both models are identical, including the context size of one million tokens, output price of 0.75$ per million tokens, and 300 tokens per second output speed. However, 3.8 Flash does have a reduction in processing by 50% for batch treatment.

The model's improvements include a more intelligent approach, with more reasoning steps, smaller units, iterative tool calls, and in-process verification. Benchmarks show 3.8 Flash outperforming 3.7 Flash by 2.4% in Agent Financial Vals v2, 1.3% in HLE-Verified, and 1.2% in Harvey Legal Agent tests. While the exact scores for SWE-Bench Pro, Terminal-Bench, and OSWorld are not published, they are expected to be close to the 3.7 Flash results.

In terms of cost, 3.8 Flash in high mode costs about 45% more per task than 3.7 Flash, with an additional 14% in real-time. However, the exact impact on your specific use case would depend on the type and volume of tasks you're handling. Google has not provided a comprehensive cost analysis yet.