Google releases Gemini 3.6 Flash and starts training next-gen Gemini 4

Google has released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, while also confirming that training has started on Gemini 4, according to the company's blog.

Gemini 3.6 Flash replaces the previous 3.5 Flash model. According to Google's benchmarks, the new version uses 17% fewer output tokens based on Artificial Analysis metrics, with savings reaching 65% on some tests, including DeepSWE. The model is also cheaper, costing $1.50 per million input tokens and $7.50 per million output tokens.

For coding tasks, Gemini 3.6 Flash showed improved results, generating more production-ready code. Its score on DeepSWE increased from 37% to 49%, while performance on MLE Bench, a test focused on machine learning research tasks, grew from 49.7% to 63.9%. On GDPval-AA v2, which evaluates office-related tasks, the model scored 1421 points compared with 1349 for its predecessor. Performance on OSWorld-Verified, a computer-use benchmark, improved from 78.4% to 83%. Computer control is now available as a built-in tool in the Gemini API and Gemini Enterprise.

Google also highlighted improvements in safety. Gemini 3.6 Flash includes stronger protection against jailbreak attempts in sensitive areas, including chemical, biological, radiological, and nuclear (CBRN) risks, as well as cyberattacks. At the same time, the company says the model is less likely to refuse legitimate requests.

• Alongside Flash, Google introduced Gemini 3.5 Flash-Lite — the fastest model in the lineup, reaching speeds of 350 tokens per second. It costs $0.30 per million input tokens and $2.50 per million output tokens. The model outperforms Gemini 3.1 Flash-Lite and even beats Gemini 3 Flash on benchmarks such as SWE-Bench Pro (54.2% vs. 49.6%) and OSWorld-Verified (74% vs. 65.1%).
• Google also unveiled Gemini 3.5 Flash Cyber, a specialized model designed to find and fix code vulnerabilities together with the CodeMender agent. Due to the potential for dual use, access is currently limited to governments and trusted partners through a pilot program.

Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are already available in the Gemini app, as well as for developers through Google AI Studio, Android Studio, and Gemini Enterprise. Gemini 3.5 Pro remains in testing with partners, with no public release date announced.

Google's AI team has also started what it calls its most ambitious pretraining effort ever for Gemini 4. The company has not yet revealed details about the next-generation model.

Gemini 3.5 Flash was first introduced at Google I/O in May. Read our full coverage on the model on our website: