Sber introduced GigaChat 3.5 Reasoning on 10 September 2026, a model aimed at complex software, analytics and business-operation tasks. It pairs reasoning-focused training with an extended context window so it can plan a piece of work, carry it out and then check its own output.
The release fits a clear 2026 trend: models tuned less for quick chat answers and more for multi-step work, where the ability to plan and verify matters more than raw speed.
What it means: 'reasoning' models are built for the kind of task a business actually wants delegated, working through a whole process rather than answering one question. The trade-off is that they think longer and cost more per task, so they pay off when the task is worth the wait.