GPT-5.2 lands, MCP moves to a foundation, Devstral 2 and GLM-4.6V

What shipped

OpenAI

GPT-5.2 system card update

GPT-5.2 Instant, Thinking and Pro: OpenAI's response to Gemini 3, aimed at professional work, long context and tool use.

Anthropic

Donating the Model Context Protocol and establishing the Agentic AI Foundation

MCP, goose and AGENTS.md move under a Linux Foundation fund so no single lab controls the plumbing.

Mistral AI

Devstral 2 and Mistral Vibe CLI

A 123B open-weight coding model, a 24B small one, and a terminal agent to drive them.

Z.ai

GLM-4.6V

An open vision-language model that can call tools using images directly, in 106B and 9B sizes.

Runway

Runway GWM-1

A general world model with variants for explorable worlds, robotics simulation and avatars.

Allen Institute for AI

Olmo 3.1 32B Think

Extended reinforcement-learning training on top of the fully open Olmo 3 recipe.

The week's model launch was OpenAI's GPT-5.2, but the item that changes how I build is a governance announcement.

MCP gets a neutral home

Anthropic donated the Model Context Protocol to a new Agentic AI Foundation under the Linux Foundation. OpenAI contributed AGENTS.md and Block contributed its goose agent, with other companies joining as members. MCP is the way most agents now plug into outside tools and data, and until now one lab controlled it.

If you run a company on agents, you write connectors: to your mail, your drive, your supplier database, a customs system. Those connectors are an asset, and you don't want them stranded if a vendor changes course. A neutral foundation lowers that risk. It doesn't remove it, since standards drift, but the direction is right, and it reads as the labs agreeing that interoperability matters more than lock-in.

GPT-5.2

OpenAI released GPT-5.2 on Thursday, in Instant, Thinking and Pro versions, pitched at professional tasks like spreadsheets, presentations, long documents and tool use. I link the system card update because OpenAI's launch page blocked my link check. It arrives about three weeks after Gemini 3 and Opus 4.5, which tells you how tight the race is. With three labs trading the lead every few weeks, I plan for model-swapping instead of loyalty. That's also why I keep prompts and procedures in plain files I own, so a switch is a configuration change.

Open weights and world models

Mistral released Devstral 2, a 123B coding model, and a 24B small model, along with Mistral Vibe, a command-line coding agent. Z.ai released GLM-4.6V, a vision-language model with native tool calling, so a screenshot can go straight into a tool call. The Allen Institute followed Olmo 3 with Olmo 3.1, further trained with reinforcement learning on the same fully open recipe.

Runway announced GWM-1, a world model family with variants for explorable environments, robotics simulation and conversational avatars. World models remain more promise than product, but this is a company shipping something users can try.

What I'd watch next

Gemini 3 launched as a Pro model, so the obvious next step is a faster, cheaper Flash version. I'd expect Google to release one before the holidays.