Sber has introduced GigaChat 3.5 Reasoning – a new version of its flagship neural network with a reasoning mode. Instead of an instant answer, the model can first break down a task into several steps, create a plan, check intermediate results, and, if necessary, change the course of the solution.

This mode is designed for tasks where a single query is not enough. For example, the neural network can search for contradictions in contracts, perform financial calculations, debug code, or consistently plan a trip considering budget, tickets, and hotel. The user decides when to enable reasoning with a separate button.

According to Sber, the new model has significantly improved in programming, planning, and working with external tools. In some tests, the growth was substantial: for example, in Live Code Bench v6, the result increased from 56 to 85 points, and in Natural Plan, from 64 to 80. At the same time, the company claims that for mathematical tasks, the model uses an average of 37% fewer tokens than DeepSeek V4 Flash Preview.

GigaChat 3.5 Reasoning is already available to GigaChat users. Developers can also use the open model for free under the MIT license and embed it into their products. Business access via API – a software interface for connecting the model to other services – is promised to be opened later.

Sber calls GigaChat the only Russian open model with such a reasoning mode. The new version is based on GigaChat 3.5 Ultra, which was additionally trained to solve mathematical and programming tasks step-by-step and independently determine when to check a previous step or refer to an additional tool.