Russia’s Sber Launches GigaChat 3.5 Reasoning to Take on Complex AI Tasks

Sber has unveiled GigaChat 3.5 Reasoning, a new flagship model that can break complex problems into sequential steps, use external tools, check intermediate results and revise its approach before producing an answer.

The Russian bank said GigaChat 3.5 Reasoning is the only Russian open-source model with a reasoning mode. The model is already available to GigaChat users and developers, with developers able to integrate it into their products for free under the MIT licence. Sber said an API version for businesses will be available shortly.

The company says that unlike a conventional model that attempts to answer a question immediately, GigaChat 3.5 Reasoning can first create a plan, identify missing information, perform multiple actions, verify its intermediate conclusions and correct errors before returning the final response.

The reasoning mode is designed for tasks where a single answer or search is not enough, including finding inconsistencies between contract clauses, carrying out financial calculations and debugging complicated code.

“From now on, GigaChat can independently break down a task into stages, identify missing data, validate intermediate conclusions, and modify the solution if necessary,” said Anton Frolov, Senior Vice President and Head of GenAI Development at Sberbank. “This approach is crucial in agent-based scenarios, programming, and analytics — whenever the outcome depends on a whole sequence of decisions. This is a vital step toward AI that can independently build a path to a result, act proactively, and interact with other services. GigaChat is the only Russian model with a reasoning mode.” 

The new version is particularly aimed at coding and agent-based tasks that require several actions to happen in sequence. Sber gives trip planning as one example. GigaChat can ask for missing details, select a flight and hotel, check whether the proposed itinerary fits the budget and then restructure the plan if it finds a discrepancy.

In following complex instructions, planning and code generation, GigaChat 3.5 Ultra Reasoning has approached the performance of leading global models, according to Sber. Its IFBench score increased from 44 to 77 points, while its Natural Plan score rose from 64 to 80. On Live Code Bench v6, its score increased from 56 to 85 compared with the version without reasoning.

The model is also designed to be more economical in its use of reasoning tokens. Sber said GigaChat 3.5 Reasoning uses significantly fewer tokens than leading open-source models with comparable answer quality. 

In mathematical problem-solving, it uses an average of 37% fewer reasoning tokens than DeepSeek V4 Flash Preview.

Sber expects the reasoning model to be used across finance and banking for scenario forecasting and checking transactions against regulatory requirements. In legal services, it can search for inconsistencies in contracts and check documents against regulations or precedents.

Sber built GigaChat 3.5 Reasoning on top of GigaChat 3.5 Ultra, specifically training it to solve problems through sequential reasoning rather than simply producing an immediate response.

During training, the model was presented with mathematical and coding problems and taught to approach them through different solution paths. Automatic verification was then used to determine which reasoning approaches produced correct answers, with those paths reinforced during training.

The process was intended to teach the model not just how to arrive at an answer, but how to construct a sequence of actions. It also learned to decide when it should use an external tool and when it should return to an earlier step and revise its approach.

Under the hood, GigaChat 3.5 Reasoning uses Sber’s proprietary architecture with linear attention technology, which the company says makes the model more efficient when working with long contexts.