All insights

Quarterly review30 September 20251 min read

GPT-5, an Olympiad milestone and the first EU rules for model makers

Models reached gold-medal standard at the International Mathematical Olympiad, OpenAI released GPT-5 and open-weight models, and EU obligations for model providers began.

What happened this quarter

In July, experimental models from Google DeepMind and OpenAI both reached gold-medal standard at the International Mathematical Olympiad, writing proofs in natural language. On 2 August, the EU AI Act's obligations for providers of general-purpose AI models began to apply.

In early August, OpenAI released gpt-oss, its first open-weight models in several years, and Anthropic released Claude Opus 4.1. On 7 August, OpenAI launched GPT-5, which decides for itself how long to think about a question. Google DeepMind showed Genie 3, which generates interactive 3D worlds from text. On 29 September, Anthropic released Claude Sonnet 4.5, with a focus on long-running agent work.

Our take

The Olympiad result is a milestone worth pausing on. Writing a rigorous proof is not pattern matching; it requires a chain of reasoning that holds up to an expert's scrutiny. Two years ago, models struggled with school maths.

For the organisations we work with, the practical lesson is about evaluation. When models change this quickly, last year's view of what AI can and cannot do is out of date. The answer is not to chase every release, but to keep a set of tests built from your own work, and to rerun them when a new model arrives. That way you can see, with evidence, whether a new model is better for your task.

Open-weight models from a major lab also widen the options for sensitive work, where data must stay on your own infrastructure.

Keep reading

Related insights