October AI news

October was mostly about open-source. Alongside the new models, we saw a shift in where models run. Instead of just faster LLMs, an ecosystem is emerging where models work inside the tools you use every day. And while capabilities keep growing, the active part of the models keeps shrinking (active parameters in MoE). Performance is no longer just about parameter count, but mostly about data, training, and architecture.

A small but capable open-source model from ServiceNow. It scores around 52 on the Artificial Analysis Intelligence Index and holds its own against much larger models (DeepSeek R1 0528, Gemini-Flash, and the like). It is a good example that data quality and training method now often beat raw parameter counts.

Useful resources: Artificial Analysis

OpenAI introduced "apps" for ChatGPT and the Apps SDK. In practice, these are MCP servers that provide custom UI directly inside ChatGPT, which makes third-party integration into GPT conversations far more user-friendly. It also opens a whole new, large marketplace for building apps, reaching more than 800 million active users.

Useful resources: OpenAI

MiniMax published M2. It reached a new high among open-source models in the Artificial Analysis ranking (around 61 points) while using about 10 B active parameters (out of 200 B total in MoE). It is currently the smartest open-source model and the fifth smartest model overall. M2 targets agents and code first.

Useful resources: Artificial Analysis

Cursor 2.0 moves the editor toward an agent-first approach. In the Agents view you can run up to 8 agents in parallel (with different models if you want), compare their strategies in one overview, and pick the best result. With git worktrees, changes run in isolated repo copies, so you can keep coding normally in the same project without collisions. The proposals stay clearly separated, which makes code review and auditing easier. Cursor 2.0 also ships Composer 1, its own model for fast agent loops (the team claims up to 4x the speed of similarly capable models). In practice, this means fewer dead ends, more usable solution variants, and faster iteration during refactoring and exploration, because Cursor tries approaches in parallel.

Useful resources: Cursor

October shows nicely that parameter count is no longer everything. Data quality, architecture, and training method matter more and more. The gap between open-source and closed-source models keeps shrinking (the fifth best-performing model is open source), so teams have more room to optimise price and deployment.

At the same time, the active size of models keeps dropping even as their capabilities grow. These are increasingly fragmented MoE approaches that cleverly switch on only the expert parts they need. And the most important shift: value is moving from chasing new models to putting AI to work, into IDEs, chat, and internal tools, where it actually makes people faster and lifts the quality of their output.