What happened
On 13 July 2026 Sber (Russia's largest bank and a major domestic AI developer) introduced GigaChat 3.5 Ultra, its new flagship AI model. The model brings a set of substantial upgrades over previous versions: it generates long texts up to 4x faster, consumes fewer resources and is nearly half the size. GigaChat 3.5 Ultra performs noticeably better on code writing, math, processing large volumes of text and on scenarios that involve autonomous agents. Under the hood sits Sber's own in-house architecture built around linear attention.
The updated model is broadly available. Inside the GigaChat AI assistant it is free for anyone who wants to use AI for personal or work tasks. On top of that, GigaChat 3.5 Ultra ships as Open Source for developers worldwide, so they can embed it into their own services and build their own AI agents. On benchmarks that measure programming ability, math, complex multi-step task execution and the quality of Russian-language dialogue, GigaChat 3.5 Ultra beat Sber's previous flagship and came close to strong open models such as DeepSeek 3.2 while being far more compact.
Why it matters
The GigaChat 3.5 Ultra release underlines the growing potential of domestic Russian AI development. The model's improved handling of code and long texts, along with its efficiency in autonomous scenarios, makes it competitive on the global market. Models like this push digital transformation across industries: they let teams automate complex processes, raise productivity and build new products and services. Opening the model up to developers fuels innovation and helps an ecosystem of AI solutions take shape.
What it means for vibe coding
For anyone practicing vibe coding, GigaChat 3.5 Ultra is a serious tool. The model's stronger code generation and code review skills, plus its accuracy on math and finance problems, make it a valuable development assistant. Being able to work through long texts such as contracts or technical documentation without losing context simplifies handling large bodies of information. Its autonomous capabilities cover routine work — monitoring, data processing and scheduled report generation — freeing up developer time for harder, more creative problems. Open access to the model also opens the door to experiments and to building custom AI agents for specific vibe-coding needs.