All insights

Weekly25 September 20261 min read

This week in AI: new flagships, released carefully

Anthropic released Claude Opus 5.5 and Google introduced Gemini 4 Argon, which went first to trusted cyber defenders. Staged releases are becoming normal for the most capable models.

What happened

On 22 September, Anthropic released Claude Opus 5.5. Google introduced Gemini 4 Argon, a frontier model with advanced reasoning, and made it available first to trusted cyber defenders through its Fairwind programme before a wider release.

Our take

The way Gemini 4 Argon was released is worth noting. Giving defenders early access means security teams can find and fix weaknesses before the same capabilities are widely available. Other labs have taken similar approaches with cyber verification schemes. As models get better at finding software vulnerabilities, we expect staged releases like this to become normal.

For most organisations this means a short wait before the newest models are generally available, and that is no bad thing. It gives time for independent testing and for early problems to surface.

Our own approach is cautious by design. We test new models against our own tasks as soon as we can, but we only use them in client work once they are generally available, covered by the provider's enterprise terms, and showing better results on the work in front of us. A new flagship is good news. It still has to earn its place.

Keep reading

Related insights