Gemini 3.8 Flash Strengthens Long-Horizon Coding Agents, but Migration Requires Sampling and Thinking Parameter Changes

Google has made `gemini-3.8-flash` generally available (GA), expanding its positioning beyond a low-latency chat model toward long-horizon software engineering, tool orchestration, and enterprise knowledge work. The model accepts text, images, audio, video, and PDFs, with a 1 million-token input window and up to 64K tokens of text output. It also supports function calling, search, and computer use. Developers can choose among `low`, `medium`, and `high` thinking levels to balance quality, latency, and token consumption. The default is `medium`; `minimal` is not supported.

Google has made 3.8 Flash the default underlying model for Antigravity managed agents and its SDK, indicating that this update focuses on multi-step execution rather than merely improving single-turn question-answering scores. The company says the model achieves a success rate above 70% on the DeepSWE v1.1 long-horizon software engineering benchmark, 54.9% on HLE-Verified, and 61.4% on Vals Finance Agent v2. However, scores for different models may use figures published by their respective providers. Google also says results from the updated evaluation process may not be directly comparable with figures in older model cards. Although its computer-use performance on OSWorld 2.0 improved over 3.7 Flash, it still trails the leading models.

API migration involves more than replacing the model name. Google requires developers to remove `temperature`, `top_p`, `top_k`, and `candidate_count`, and to replace the numeric `thinking_budget` with the string-based `thinking_level`. Multi-turn conversations should use the server-side `previous_interaction_id`, while function responses must preserve call-identification information. If an existing agent depends on fixed sampling parameters, prefilled model messages, or locally managed full conversation histories, teams should run regression tests using real tool trajectories before upgrading.

The model card also lists limitations including occasional timeouts, proactive use of more tokens on complex tasks, and knowledge cutoffs that may extend only to early 2025 in some domains. Of particular concern for Chinese-language teams, Google’s automated multilingual safety metric deteriorated by 5.4 percentage points compared with 3.7 Flash. Google says its manual review found that most of the decline resulted from false positives, but it did not provide a language-by-language breakdown. Promotional input/output pricing through the end of the year is $0.75/$3.75 per million tokens and will double in 2027, so cost assessments should not rely solely on current prices.