GigaChat: 7th in the World on a Blind Russian-Language Ranking, but 78 Elo Points Behind the Leader
Putin said two Russian models are among the best in the world, and the claim can be checked. On llmarena.ru, people blindly pick the better of two answers, and Elo scores are calculated the same way as in chess. GigaChat 3.5 scored 1121 and ranked 7th out of 112 models. The leader, Gemini 3 Pro Preview, scored 1199, which is 78 points higher. Alice AI LLM ranks 20th with 1069. For Russian-language tasks, both models are workable, but for complex agents, Western models still lead.
AI-processed from Habr AI; edited by Hamidun News
On November 18, at the APEC summit in Shenzhen, Vladimir Putin said: "Our two flagship language models are among the best in the world." This claim can be checked against the open ranking at llmarena.ru.
There, a user asks a question, two anonymous models answer, and the person votes for the better answer before seeing the model names. Votes are converted into Elo scores, the same system used for chess players: the more often you win, the higher your score. In the November 2026 snapshot, GigaChat 3.
5 is 7th in the world with 1121 points, while the leader, Gemini 3 Pro Preview, scored 1199. The gap is 78 points. In practice, this means GigaChat is a workable option for Russian-language tasks, while for complex agents it is worth testing Western models first.
The author of the Habr article is not an ML engineer but an ordinary user who needs answers in Russian without a VPN. He collected data on all four Russian model families and compared it with the developers' claims.
How the ranking works and what it can be trusted for
LLM Arena works like a blind tasting. The final score is calculated from millions of head-to-head duels using the Bradley-Terry model, the same mathematics behind the chess Elo rating. The ranking covers 112 models, and the author's snapshot includes votes up to November 9, 2026.
The method has limits. It shows which answer people preferred more often, but it does not show which model will solve an Olympiad problem or build a complex project. On the other hand, developers cannot tune results to a benchmark. And because the ranking is live, the numbers must be read together with the snapshot date: in September, Alice was 19th with the same 1069 points.
Four families: who is who
GigaChat (Sber). The flagship is GigaChat 3.5 Ultra, released in July 2026. It uses a MoE architecture with 432 billion parameters (28 billion active) and a context window of 262 thousand tokens. The weights are open under the MIT license, so the model can be downloaded, fine-tuned, and run on your own hardware. In the ranking, it holds 7th place with 1121 points. Nearby are GPT-5.2 with 1122 and DeepSeek V3.2 with 1119, so on Russian-language tasks GigaChat performs almost on par with them.
Alice AI LLM (Yandex). The flagship has 235 billion parameters (MoE) and was trained from scratch on Russian-language data. The model is closed, but it is deeply integrated into services: search, the browser, smart speakers, and, through the API, it can book a table or place an order with Taxi and Eda. On September 21, Yandex released the pretrained Alice AI Foundation model with 80 billion parameters for commercial use. On the leaderboard, Alice is in 20th place with 1069 points, alongside Gemma 3 27B and Claude Sonnet 4.
T-Pro and T-Lite (T-Technologies). The Gen-T family is built on Qwen and fine-tuned on a Russian corpus. According to the developer, this cut costs by roughly 90% compared with training from scratch. The current T-pro-it-2.0 and 2.1 run on Qwen 3 with 32 billion parameters and are tuned for instruction following and tool calling. The lightweight T-Lite with 7 billion parameters is designed for fast chatbots. The weights are open under Apache 2.0. In the September snapshot, T-pro-it-2.0 ranks 36th with 1024 points.
Cotype (MWS, MTS ecosystem). This is not a general-purpose assistant but a set of business models trained on corporate correspondence and documents. Pro 3, with 27 billion parameters, supports a 262-thousand-token context, roughly the length of a full annual report. Light 3, with 9 billion parameters, is intended for routine workflows. The multimodal VL model, with 32 billion parameters, reads handwritten text from scans, and the open Nano model, with 1.5 billion parameters, suits drafting. In the independent MERA benchmark, Pro 3 placed third. Cotype is not on the public llmarena.ru leaderboard, so it cannot be compared with the others on this scale.
The bottom line
The leader in this snapshot is Gemini 3 Pro Preview with 1199 points, followed by Qwen3.6 Max (1165), GPT-5.4 (1161), and Claude Opus 4.7 (1149). GigaChat trails the top spot by 78 points, Alice by 130, and T-pro by 175. In other words, only one Russian model makes the global top ten on this ranking, not two, and the second lands in the top twenty.
The author places both flagship models in a second tier. World leaders, as well as the Chinese DeepSeek and Qwen, are noticeably stronger in long agentic chains, complex programming, and multi-step reasoning. However, for Russian-language tasks such as writing, translation, or summarizing a document, the gap is barely noticeable in practice. Add to that the open weights of GigaChat, T-Pro, and Alice Foundation, which Western flagships do not offer, along with integration into banking, search, and marketplaces.
The practical takeaway for developers and product managers is this. If the task is in Russian and local deployment or in-country data matters, GigaChat and T-Pro are solid options. If you are building a complex coding agent, first test Gemini, GPT, or Claude on your own tasks. The evaluation has one main limitation: this is a preference ranking for Russian, not a universal exam, and positions may shift within a month.
- Nov 22, 2026A Backend Developer With No Unity Experience Built a 3D Game With an AI Agent: Subscriptions Cost $112
- Nov 22, 2026A Solo Researcher Spent 1.5 Years Trying to Prove the Riemann Hypothesis with Claude and Lean, and Failed
- Nov 22, 2026OpenAI's Chief Scientist Proposes an AI Speed Limit: GPT-4 Perplexity at 0.91, Now Below 0.8
- Nov 22, 2026DeepSeek Veteran Open-Sources Linger: A Local AI Agent That Responds Up to 5x Faster
Key Questions Answered
How the ranking works and what it can be trusted for?
LLM Arena works like a blind tasting. The final score is calculated from millions of head-to-head duels using the Bradley-Terry model, the same mathematics behind the chess Elo rating. The ranking covers 112 models, and the author's snapshot includes votes up to November 9, 2026. The method has limits. It shows which answer people preferred more often, but it does not show which model will solve an Olympiad problem or build a complex project. On the other hand, developers cannot tune results to a benchmark.
What are the key details on Four families: who is who?
GigaChat (Sber). The flagship is GigaChat 3.5 Ultra, released in July 2026. It uses a MoE architecture with 432 billion parameters (28 billion active) and a context window of 262 thousand tokens. The weights are open under the MIT license, so the model can be downloaded, fine-tuned, and run on your own hardware. In the ranking, it holds 7th place with 1121 points. Nearby are GPT-5.2 with 1122 and DeepSeek V3.2 with 1119, so on Russian-language tasks GigaChat performs almost on par with them. Alice AI LLM (Yandex).
What are the key details on The bottom line?
The leader in this snapshot is Gemini 3 Pro Preview with 1199 points, followed by Qwen3.6 Max (1165), GPT-5.4 (1161), and Claude Opus 4.7 (1149). GigaChat trails the top spot by 78 points, Alice by 130, and T-pro by 175. In other words, only one Russian model makes the global top ten on this ranking, not two, and the second lands in the top twenty. The author places both flagship models in a second tier.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.