Model updates: September 2026

Two of the three big providers replaced their headline models in September, and the pattern of the year held: the new mid-priced model does most of what the expensive one does. The newer pattern is a gate in front of the most capable versions.

What came out

• Anthropic released Claude Fable 5.1, its most capable model, on 1 September, and Claude Opus 5.5 on 22 September. Anthropic says Opus 5.5 “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5”. The same day it raised the five-hour usage limits on its Pro, Max, Team and seat-based Enterprise plans. Sonnet 5.5 and Haiku 5.5 are promised “in the coming weeks”. What switching to Opus changes has its own page.
• OpenAI moved to GPT-6. Microsoft’s Azure listings date GPT-6 Astra to 3 September and GPT-6 Sol and Luna to 22 September, all three with a context window, the amount of text a model holds in mind at once, of just over a million tokens. Which ChatGPT plans get which one is set by OpenAI and listed in its own release notes.
• Google’s newest is Gemini 3.8 Flash, which Google’s own developer tools started supporting in September. Google’s Gemini changelog has the detail.
• Microsoft put two of these into Copilot within days: Claude Fable 5.1 in Copilot Cowork and Copilot Studio on 1 September, and GPT-6 Astra in the same two on 4 September. Its roadmap did not list Opus 5.5 as of 26 September.

The gate

The top of the range is no longer sold to everyone who pays. Anthropic released a second version of Fable 5.1, Claude Mythos 5.1, with safeguards built for cybersecurity and life-sciences work, and it is “only available to a set of US organizations” vetted through Anthropic’s access programmes. Microsoft’s documentation says GPT-6 Astra may switch on stricter safety controls when its safety systems spot elevated risk. Both providers are holding back part of what their strongest models can do, one by vetting who gets it and the other by tightening controls when the risk looks high. For most readers nothing changes. If your organisation does security testing, the strongest model may now need an application rather than a subscription.

The most capable model is no longer the default choice. Anthropic’s own documentation now tells developers to start with Opus 5.5 and move to Fable 5.1 only when Opus falls short.

Also worth knowing

Anthropic says text from its new models now carries a watermark, a statistical pattern in word choice that shows how likely it is that Claude wrote a passage, with older models to follow “over the coming months”. Only “eligible organizations” such as regulators, the media and fact-checkers can run the detector for now. Anthropic says it is doing this “to comply with the EU AI Act”, and that “a complete rewrite where every word is replaced” removes it. If you are deciding whether to say you used AI, the watermark is a reason to be upfront rather than a reason to rewrite.

What this means in practice

If you pay for Claude, the change you will notice is the usage limit, not the model name. If you use Copilot at work, the new models reach you only once your organisation switches them on, and in September they started in Cowork. For how the assistants compare on real tasks, see the writing and coding comparisons, and for last edition’s picture, July’s model updates.