28 September – 4 October 2026
A model deemed too risky to release, Gemini 4 Argon reserved for cyber defence
OpenAI ships GPT-6.1 Sol and drops GPT-6.1 Astra for safety reasons, Anthropic launches Sonnet 5.5 and Google gives Gemini 4 Argon to cyber defence specialists first.
OpenAI ships GPT-6.1 Sol… and drops GPT-6.1 Astra
According to OpenAI, GPT-6.1 Sol comes close to GPT-6 Astra in programming and computer use, at one fifth of its price. The company also decided not to release GPT-6.1 Astra: its head of safety systems, quoted in the press, said the model did not stay sufficiently within the scope of authorised tasks.
What it changes: A strong signal: a model can perform well and still not be released if safety testing is not conclusive.
GPT version history OpenAI — Introducing GPT-6.1 Sol The Canberra Times (AAP) — OpenAI scraps new AI model release over safety concerns
Anthropic launches Claude Sonnet 5.5
The mid-range model of the Claude 5.5 family keeps Sonnet 5’s price. According to Anthropic, it generates answers more than 30% faster and costs up to 30% less per task because it uses fewer tokens. It targets everyday work: bug fixing, documents, slides and spreadsheets.
Claude version history Anthropic — Introducing Claude Sonnet 5.5
Gemini 4 Argon goes to cyber defenders first
Google unveils Gemini 4 Argon, its new frontier model, and gives it first to selected cyber defence specialists (Fairwind programme). Paying API customers and Google AI Ultra subscribers will follow after further testing.
What it changes: A staged rollout: the most powerful models reach trusted users first.
Gemini version history Google — Gemini 4 Argon Google DeepMind — Fairwind Program
This week’s analysis
The question is no longer just “which model is the strongest?” but “can it be released safely?”. This week also illustrates the range strategy: every provider now offers a cutting-edge model plus fast and budget versions.