Google has announced the Gemini 3.1 Pro AI model and opened it to early access for developers, businesses, and consumers. The company claims it will improve its ability to perform complex analysis, multi-step planning, and solve atypical problems.
Gemini 3.1 Pro is an update to the flagship Gemini 3 model, introduced in November 2025. According to Google, this model is the basis for the recently updated Deep Think system.
On the Humanity’s Last Exam, a test that measures knowledge in complex subject areas, Gemini 3.1 Pro scored 44.4%. In comparison, Gemini 3 Pro scored 37.5% and GPT-5.2 scored 34.5%.
Google also notes progress on the ARC-AGI-2 benchmark, which tests the ability of models to solve logic problems that cannot be directly learned during training. Gemini 3 scored 31.1% on this test, while the new version achieved 77.1%.
However, Gemini 3.1 Pro did not top the Arena (formerly LM Arena) ranking, where users rate models based on the quality of their answers. In word processing tasks, Claude Opus 4.6 is ahead, while in programming tasks, Opus 4.6, Opus 4.5, and GPT-5.2 High remain ahead.
Google demonstrates the model’s improvements using the example of graphics generation and simulations. In particular, Gemini 3.1 Pro is able to create complex SVG images directly from text queries. For developers of agent systems, the model also showed better results – its score in the APEX-Agents test is almost twice as high as in the previous version.
The model is already available in preview via AI Studio and Antigravity. Enterprise customers can use it in Vertex AI and Gemini Enterprise. For regular users, Gemini 3.1 Pro is available in the Gemini app and NotebookLM service.
The API cost for developers remains unchanged at $2 per million input tokens and $12 per million output tokens. The size of the context window has also not changed and is 1 million tokens per input and 64,000 per response.
Read also: Google introduces music generation capability to Gemini
Share:
Посилання скопійовано