Oct 5, 2026

A gated frontier model, OpenAI's always-on agent, and the EU's rules in force

Google's newest model reaches security defenders before anyone else, OpenAI shows an 'always-on' agent while shelving a model over safety concerns, and the EU's AI transparency rules are now enforceable.

1. 01Models

   Google announces Gemini 4 Argon — to cyber defenders first ↗

   Announced on 30 September, Argon is rolling out first to security defenders through Google's Fairwind program, followed by paid API customers and Google AI Ultra subscribers; no wider date was given. Introductory pricing is $2 per million input tokens and $10 per million output tokens, and the output limit rises to 1 million tokens (from 64,000). The benchmark numbers (77.9% on DeepSWE v1.1, 68% on CWE-bench) are Google's own, as reported by Axios.

   Why it matters: Gating a model to defenders first is a new release pattern for security-capable models. And vendor benchmarks are a starting point: The Next Web notes Bloomberg reported some Google staff say it does better on benchmarks than in real use, which Google disputes. Test on your own tasks before switching.

2. 02Business

   OpenAI shows 'Dots' always-on agents at DevDay — and shelves a model over safety ↗

   At its developer conference on 29 September, OpenAI introduced Dots, described by Sam Altman as AI helpers 'that always has your back', plus GPT-6.1 Sol and an 'ultrafast' premium speed tier. The AP reports OpenAI halted the rollout of a more advanced model after researchers raised safety concerns; the model was not named, and Altman said in Q&A the company is investing more in safety, security and monitoring of AI agents. The AP article gave no pricing or availability dates.

   Why it matters: 'Always-on' agents act continuously, so the controls around them matter more: scoped permissions, approval for sensitive actions and audit logs. If you build agents, this is the design bar your users will now compare you to.

3. 03Policy

   EU AI Act: transparency rules apply, high-risk rules pushed to December 2027 ↗

   Since 2 August 2026 the Article 50 transparency obligations are enforceable: people must be told when they are interacting with an AI system such as a chatbot or voice assistant, and synthetic audio, image, video and text must carry machine-readable marking. Providers of generative AI systems already on the market have until 2 December 2026 for the marking and detection part. Separately, the Digital Omnibus (Regulation (EU) 2026/1744, in force 27 July 2026) postponed the high-risk obligations from 2 August 2026 to 2 December 2027, per DLA Piper.

   Why it matters: If you ship a chatbot or agent to EU users, disclose that it is AI now. The high-risk delay buys time, but it is a deadline moved, not removed. Get legal advice for your situation.

4. 04Models

   Open 'decision models' for agents: Cloudflare's Clef and AWS's Strands Decider ↗

   On 1 October Cloudflare released Clef (27B) and Clef-flash (9B), open-weight 'decision models' under Apache 2.0, built on Qwen bases, that answer bounded questions (yes/no, multiple choice, rankings) with typed probabilities instead of generated text; Cloudflare's post and The Register describe them. The same day AWS Strands Labs released Strands Decider 2B, an open-source model of about 2 billion parameters adapted from Qwen3.5-2B that picks among fixed options, with weights, training data and scripts published, per SiliconANGLE and MarkTechPost. Speed and accuracy figures in these announcements are the vendors' own.

   Why it matters: A decision model is a small, fast model whose only job is to choose the next step — the router idea from the orchestration course, now shipped as open weights. If you build agents, it's a cheaper way to handle easy routing and tool-selection decisions; test it on your own cases before trusting the headline numbers.