The Intake — Monday, October 5, 2026

Google DeepMind released Gemini 4 Argon on September 30, 2026. It scores 77.9% on DeepSWE v1.1 and 68% on CWE-bench v1. Those scores place it ahead of GPT-6 Astra and Claude Opus 5.5 on 12 of 18 disclosed benchmarks. Supported output runs up to 1 million tokens. Introductory pricing is $2 per million input tokens and $10 per million output tokens. Google disclosed the rates will rise to $4 per million input and $20 per million output after the introductory period. No end date for the introductory period was published.

Google says its internal deployment of Argon identified a previously unknown critical vulnerability in healthcare software. The model ships with chain-of-thought monitoring on by default; Google says it stops execution when misalignment is detected. Access is currently restricted to API customers and Google AI Ultra subscribers.

If you need the guardrail-free variant for adversarial security testing, access requires applying to Google's Fairwind Program for cyber defenders.

If you generate text or code for EU users, OpenAI's textGrain watermarking will apply to your output by December 2, 2026 — and the opt-in to test it before then is live today.

OpenAI announced textGrain on October 5, 2026. OpenAI says it was co-developed with the University of Pennsylvania and Yale. The system uses entropy-calibrated watermarking. Rather than marking content visibly, it statistically influences word-choice probabilities to embed a provenance signal. OpenAI's published detection rates are approximately 80% at 200 tokens and 95% at 400 tokens. Detection drops to approximately 17% if a quarter of words are replaced with synonyms.

EU ChatGPT and Codex users receive the watermark automatically in the coming weeks. The EU AI Act compliance deadline is December 2, 2026. API users worldwide can opt in from October 5; the feature is off by default. Detector access is restricted to approved researchers and expert organizations under the EU Code of Practice on transparency. Operators cannot self-verify watermark presence.

If you're building on the API and want to understand what textGrain does to your specific content before the mandate takes effect, the opt-in path is open today. Independent verification of watermark presence is not available to you — detector access requires working with an approved researcher or relying on OpenAI's published accuracy figures.

If your agent workflow makes routing or tool-selection decisions by calling a full generation model, there's now a scoring path that handles that without generating text.

AWS Strands Labs released Strands Decider 2B on October 1, 2026. It is a 2-billion-parameter model built on Qwen3.5-2B. A pointer head of approximately 1 million parameters handles the scoring. It evaluates developer-supplied option lists and returns a selection without generating output. AWS reports median latency on an NVIDIA RTX 3090 of 106–115 milliseconds. Strands Labs reports the model briefly ranked first on JevBench for its size class. It is available under Apache 2.0 on Hugging Face and GitHub.

The model was developed by Marc Brooker, a distinguished engineer at AWS. Strands Labs positions it as a free alternative to TypeSafe's proprietary Jev decision model. Named use cases include model routing, tool selection, context management, guardrail enforcement, policy classification, and output evaluation.

If your workflows include routing calls where the answer comes from a fixed option set, Strands Decider 2B is the testable path for replacing full generation calls on that layer.

GPT-6.1 Sol, announced at OpenAI's DevDay 2026 on September 29, is priced at $2 per million input tokens and $10 per million output tokens. That is one-fifth the rate of GPT-6 Astra. Cached input tokens are $0.10 per million. The model is available via API and in Codex. Codex was extended to cloud environments for remote task initiation.

OpenAI also announced Dots — persistent agents that maintain ongoing tasks across sessions. Dots run in dedicated cloud compute environments with browser access. They connect to over 4,000 application integrations. ChatGPT Space is a shared workspace where humans and Dots agents co-edit pages with common context. It rolled out to Pro and Business subscribers on October 1.

The Decisions API entered limited preview on September 29. It uses OpenAI's Luna model to classify content and route requests from predefined option sets. Requests are handled without invoking a generation model. If you're building on OpenAI's stack and routing decisions against a fixed option set, the Decisions API is the path to test while it's in limited preview.

---