Gemini 4 Argon debuts to cyber defenders, Sonnet 5.5, GPT-6.1 Sol

Google's launch was the most unusual one this week, because the model went to a narrow audience first.

Gemini 4 Argon

On 30 September Google announced Gemini 4 Argon as its new frontier model, designed for real-world software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense. It is rolling out only to a limited group of trusted cyber defenders through the Fairwind Program while Google keeps testing safeguards before a wider release. Google plans to open it to paid API customers and Google AI Ultra subscribers at an introductory $2 per million input tokens and $10 per million output, rising later to $4 and $20.

The numbers Google reports are a 1M-token output limit (up from 64K), 77.9% on DeepSWE v1.1, first place on the Vals Index, 51.3% on Zapier's AutomationBench and 91.7% on LVBench. It also says this is the model most resilient to indirect prompt injection that it has built. All of that comes from Google's own post. I'll wait for outside tests before leaning on any of it.

Sonnet 5.5 and GPT-6.1 Sol

Anthropic released Claude Sonnet 5.5 on 28 September as the faster, lower-cost partner to Opus 5.5. List price stays at $2 per million input and $10 per million output, with output more than 30 percent faster. OpenAI announced GPT-6.1 Sol on 29 September, pitched as near-Astra intelligence at one-fifth of Astra's token prices, at $2 input and $10 output for short contexts.

Open models for agents

H company released Holo4 on 28 September, agentic models in 27B and 35B-A3B sizes for computer use. On OSWorld 2.0 the 27B scores 61.7% against 81.8% for Opus 5.5, so it is not a replacement, but the cost gap is large and the weights are open. Aleph Alpha released Kolibri 1 on 3 October, a 78B English and German reasoning model under Apache 2.0, trained from scratch in Germany. Cloudflare released Clef on 1 October, a 27B decision model.

Video, biology and talent

xAI's Grok Imagine Video 1.5 Lite generates clips of 1 to 15 seconds with audio from $0.02 a second. DeepMind introduced SynthID Bio on 30 September to watermark synthetic biology. And on 2 October Anthropic launched Claude Frontier Academy with $100 million to train 10,000 engineers, with a 12-week residency where each leads a real project at their own employer.

A paper called False Frontiers shows self-evolving search agents can learn to cheat together.

What I'd watch next

When Gemini 4 Argon opens to the public, and whether the independent numbers match Google's.