October 2026 turned out to be one of the busiest months of the year: within three weeks, the flagships of the three largest labs were all updated, the cheapest model in the Claude lineup was released, a new open model with 753 billion parameters appeared, and agent developers got official guides from OpenAI and Anthropic. We have gathered in one place everything we wrote about in the channel over the month and added the most important part — what of this makes sense for an ordinary user and what is still only for developers.
In brief: five events of the month
- Claude Sonnet 5.5 became 30% faster and cheaper at processing tasks, while nearly catching up to the flagship Opus 5.5.
- GPT-6.1 Sol from OpenAI launched with capabilities at the level of Opus 5.5, but at a lower price.
- Gemini 4 Argon from Google beat both competitors in key benchmarks, but is still available only to a limited group of testers.
- Claude Haiku 5.5 — 10 times cheaper than previous models per token and outperforms GPT-6 Luna in general tests.
- DeepSeek released a desktop application with open source code — a direct competitor to Claude Code that works with files and the terminal on your computer.
Below we break down each event in more detail, in chronological order.
Flagships: the race of three labs
Claude Sonnet 5.5 — the mid tier that nearly caught up to the flagship
At the beginning of the month, Anthropic released Claude Sonnet 5.5 — an update to the middle tier of the lineup, available immediately in the app and on the website. Key changes compared with Sonnet 5:
- operating speed increased by more than 30%;
- the cost of processing one task dropped to 30%;
- programming improved noticeably;
- on long multi-step tasks, the model outperforms GPT-6 Sol and is almost as good as Opus 5.5.
For most users, this is the most interesting news of the month: Sonnet has historically been the "default" model for those for whom Opus is too expensive, and now the quality gap between them has become minimal.
GPT-6.1 Sol — OpenAI's response a day later
Literally the next day, OpenAI introduced GPT-6.1 Sol. The formula is the same as that of its competitors this year: capabilities at the level of Claude Opus 5.5, lower price. The competition has finally turned into a cycle of "one releases a stronger model, the second lowers the price, the third prepares a response."
What this means in practice: if you pay for a ChatGPT subscription, you got a flagship-level model without changing your plan. If you are choosing between subscriptions, the difference between the top models from OpenAI and Anthropic is now in nuances, not in class — it makes sense to try both on your tasks.
Gemini 4 Argon — Google seizes the lead
The most unexpected release of the month: Google released Gemini 4 Argon, which beats GPT-6 Astra and Claude Opus 5.5 on most tests. What is known:
- output token limit — up to 1 million (the model can produce a huge document or codebase in one pass, although it sometimes works slower than competitors);
- the model is designed for complex and long tasks: programming, business analytics, cybersecurity;
- price — $2 per million input tokens and $10 per million output tokens, which is very aggressive for a flagship;
- access is still only for a limited number of testers; the official release is expected soon.
If you work with long documents or large projects, Argon is worth following first and foremost: the one-million output token limit is a qualitatively new scenario, not just "a little bit better."
Budget segment: Claude Haiku 5.5
While the flagships measure themselves against benchmarks, Anthropic released Claude Haiku 5.5 — the fastest and cheapest model of the new series. The numbers are impressive specifically against the background of the previous version:
- 39% successful solutions to complex programming tasks versus zero for Haiku 4.5;
- token cost reduced by 10 times;
- in computer control tasks — 72% versus 16% for its predecessor;
- in general tests it outperforms GPT-6 Luna, OpenAI's budget model.
Who this is for: quick answers to questions, translation, summaries, letters, Telegram bots, and simple pet projects. If you are doing something at scale via API — for example, processing thousands of messages or emails — Haiku 5.5 currently looks like the most advantageous option among closed models. You can compare it with other budget models in our AI catalog.
Open-source: GLM-5.3 without restrictions
For those who prefer to run models locally, the main news of the month is GLM-5.3. A model with 753 billion parameters and a wide context that responds quickly and handles long tasks. The developers removed almost all restrictions, and the model generates texts and instructions with virtually no blocks.
An important caveat: 753 billion parameters is not something you run on a home graphics card. Local deployment will require a multi-GPU configuration or a rented GPU server. You can estimate how much video memory is needed for a specific model and quantization in our GPU and VRAM calculator. If you do not have the hardware, GLM-5.3 is available through cloud APIs.
Tools for developers and agents
DeepSeek replaced Claude Code
In the middle of the month, DeepSeek released a desktop application with open source code for Windows and macOS. Essentially it is a competitor to Claude Code and Codex CLI, only with a graphical interface and a modular architecture:
- works with files, code, and the terminal directly on your computer;
- consists of plugins — missing functions can be added yourself;
- searches the internet, writes texts and code;
- can be assembled for your needs almost like a construction kit.
The main difference from Claude Code is the open code and the free nature of the application itself: you pay only for the tokens of the selected model.
Ready-made skills: 338 roles for one agent
At the beginning of October, developers published a set of 338 ready-made skills for AI agents — essentially descriptions of roles from CEO and CTO to lawyer, marketer, and personal assistant. An agent with such a set can draw up roadmaps, write code, analyze the market, handle SEO, calculate budgets, sort through email, and do reviews. This is a good starter kit if you want to try agents but do not know which prompt to start with.
Official guides from OpenAI and Anthropic
The two most useful publications of the month are not models, but instructions.
OpenAI released an official guide to prompts and task allocation between the Sol, Luna, and Astra models: how to work with context, how to automate routine work, how to configure reasoning depth, and how to save tokens. If you use ChatGPT every day but still write requests "by eye," start with it. And we covered the basic rules in the article how to write a prompt for a neural network.
Anthropic presented a guide to building AI agents that can be read in 9 minutes and contains a ready-made code template. An agent based on this template can daily monitor GitHub and Slack, track changes in projects, independently close tasks, send reports, and remember context while taking token limits into account. If you have long wanted to delegate part of your work routine to AI, this is the shortest path.
Additionally, a free training course on Claude Code was released — from the basics and prompts to working with Git, subagents, custom skills, and plugins, with practical assignments and review.
Claude Dashboards and Claude Motion
Anthropic is also testing two new features directly in the chat: Dashboards generates an interactive analytics panel from a project description that updates in real time, and Motion turns charts and reports into animated visualizations with live data editing. For now this is a beta for paid subscribers, but the direction is clear: chat is ceasing to be text only.
What to try right now
If you have 15 minutes and want to get value from the October releases, here is a short plan by task type:
- You write texts, translate, sort through email — try Claude Haiku 5.5: fast, cheap, and the quality is more than sufficient.
- You program — Claude Sonnet 5.5 has become noticeably better at code and faster; an open-source alternative is the DeepSeek desktop application.
- You work with large documents — keep an eye on the release of Gemini 4 Argon; the one-million output token limit changes the rules of the game.
- You want to automate routine work — start with Anthropic's 9-minute guide to agents and the set of 338 skills.
- You do not know where to start — go through our AI selection wizard: it will ask a few questions and suggest a model for your task and budget.
What to expect in November
Judging by the pace of October, in November we should expect the official release of Gemini 4 Argon, OpenAI's response to Claude Haiku 5.5 in the budget segment, and new open models — Chinese labs release them almost every month. We cover all releases in the channel on the day they come out, and at the beginning of December we will put together the next digest.