5 – 11 October 2026
A very large European open model and smaller, cheaper models
Mistral opens the Large 4 preview, Anthropic launches a cheaper Claude Haiku 5.5, OpenAI publishes 722 maths papers produced by an internal model and tests a new ad format, and Anthropic updates its usage rules.
On video
Mistral AI opens the Large 4 preview
Mistral presents Large 4, a multimodal model of about 1 trillion parameters, trained in its own data centres in Europe. The preview is available through Mistral’s API; the company plans to publish the weights (the downloadable model) by the end of the month.
What it changes: If the weights are published as planned, it will be one of the largest models anyone can download and run on their own servers.
Mistral version history Mistral AI — Introducing Mistral Large 4
Anthropic launches Claude Haiku 5.5
The smallest model of the Claude 5.5 family costs $0.10 per million input tokens and $0.50 per million output tokens for requests under 100,000 tokens. Anthropic designs it for simple, repeated tasks: summarising, sorting, or assisting a larger model. It is offered to all Claude users, including the free plan.
What it changes: For high-volume uses, the cost of a capable AI keeps falling.
Claude version history Anthropic — Introducing Claude Haiku 5.5
OpenAI publishes 722 maths papers produced by an internal model
OpenAI released 722 mathematics manuscripts produced by a model that is not yet public. Only some of the results come with a computer-checkable proof (in the Lean language), and OpenAI warns that the others may contain errors. Mathematicians still have to review them.
What it changes: An AI can propose research results, but validating them remains human work.
ChatGPT tests ads during image creation
OpenAI announces a new visual ad format, first tested in the United States from late October, during image generation for users of the free and Go plans. According to OpenAI, these ads are labelled, kept separate from the image being created, and do not influence ChatGPT’s answers.
What it changes: For now the test only covers the United States: nothing has been announced for France or Europe.
GPT version history OpenAI — Building advertising for the way people use AI
Anthropic updates its rules for using Claude
The new rules, in force from 12 November, group together the bans on deception (fake accounts, fake news sites), clarify forbidden uses related to weapons and set rules for using Claude with machines that can act physically.
This week’s analysis
The trend of the week: every provider is filling out the lower end of its range. Small, fast and cheap models such as Haiku 5.5 are becoming the “workers” that assist larger models. Meanwhile, Mistral confirms that a very large open model can come from Europe. Finally, OpenAI’s maths release is a useful reminder for everyone: a result produced by an AI must be checked before it is trusted.