# AI News — October 01, 2026: Gemini 4 Argon Locked to Vetted Partners, Rewriting Google's Own Kernel in Rust > Source: > Published: 2026-10-01 06:00:07+00:00 Good morning. Google took the spotlight today with Gemini 4 Argon, a frontier model so capable the company says only vetted cybersecurity partners can touch it for now — a release framing that drew eye-rolls on Hacker News within hours. Elsewhere, ElevenLabs doubled its valuation on a secondary, OpenAI disclosed it had shut down a model-distillation operation, and the tech industry quietly signed a self-regulation pledge with the White House that misspells its own country. **Google releases Gemini 4 Argon — but almost nobody can use it.** [Argon is Google DeepMind’s new flagship](https://deepmind.google/blog/gemini-4-argon-our-next-era-of-frontier-intelligence/), aimed at long-horizon software engineering, legal and financial analysis, and cybersecurity, with a 1M-token context window and pricing at $2/$10 per million input/output tokens. Access is restricted to “trusted cyber defenders” via Google’s Fairwind Program while the company runs voluntary pre-release review with the U.S. government, with developers and consumers waiting indefinitely. [The Verge notes](https://www.theverge.com/tech/1002980/google-gemini-4-argon) the timing — one day after OpenAI’s DevDay — and [TechCrunch reports](https://techcrunch.com/2026/09/30/google-releases-gemini-4-argon-called-its-most-powerful-model-yet/) Google’s benchmark claims put Argon ahead of GPT-6 Astra and Anthropic’s Opus and Fable on Vals’ evaluations. **What’s actually interesting is the internal use.** Buried in the announcement: Argon agents are already migrating Google’s own C/C++ codebases to Rust, including 800K+ lines of the Fuchsia OS Zircon kernel plus core libraries like re2 and libgav1. [HN commenters](https://news.ycombinator.com/item?id=49913571) seized on this as the real story, with one calling it more significant than the benchmark numbers. The frustration ran parallel: AI Ultra subscribers paying monthly still can’t access the model, and as one commenter put it, “Gemini not beating the ‘can’t release a model’ allegations.” **A community benchmark starts tracking Opus 5.5 for degradation.** [Livenerf](https://github.com/ninjahawk/livenerf) is 7 days into a 30-day pre-registered study using frozen prompts and pinned CLI versions against 78 curated “sometimes right” questions from GPQA Diamond and MMLU-Pro, with first statistically meaningful results expected October 24. [HN discussion](https://news.ycombinator.com/item?id=49901736) surfaced a competing project, BridgeBench’s Nerf Bench, which commenters claim previously detected a regression in Opus 4.6 that Anthropic later acknowledged. The split was familiar: half the thread insists post-launch degradation is real and compounding infra changes explain it, half says it’s user psychology and dopamine. **OpenAI says it disrupted a model-distillation campaign.** OpenAI published [a disclosure](https://openai.com/index/disrupting-a-coordinated-model-distillation-campaign) describing a coordinated effort to extract capabilities from its models via systematic querying — the kind of operation that trains a competing model on your outputs. Details are thin in what’s circulating, but it fits the pattern of frontier labs treating output harvesting as a security problem rather than a terms-of-service one. **ElevenLabs doubles to $22 billion on a tender.** The voice AI company closed a [$300M secondary](https://techcrunch.com/2026/09/30/ai-voice-startup-elevenlabs-doubles-valuation-to-22b/) co-led by Wellington and T. Rowe Price, letting employees sell vested equity at double February’s $11B mark. Four years old, two tenders deep — the employee-liquidity-as-retention playbook is now standard operating procedure for AI companies that don’t want to IPO yet (see also: Anthropic’s situation, Saturday). **Tech leaders sign a self-regulation pledge with Trump.** Executives from Google, Anthropic, Meta, OpenAI, xAI, and Nvidia signed a non-binding “Joint Commitment on Frontier Responsibilities,” agreeing to develop internal safety standards and meet periodically, [per The Verge](https://www.theverge.com/ai-artificial-intelligence/1002584/trump-us-ai-safety-deal-self-regulation-tech-execs). The signature block misspells “United States.” The accompanying executive order rebrands AI as “Super Intelligence” and leaves formal legislation on the table but not on the agenda. **Also worth noting.** DeepMind introduced [SynthID Bio](https://deepmind.google/blog/introducing-synthid-bio/), extending its watermarking framework to biological sequences — presumably AI-designed proteins, building on AlphaProteo. And MIT Technology Review [has an interview](https://www.technologyreview.com/2026/09/30/1145339/were-not-going-to-shoot-ourselves-in-the-foot-over-hugging-face-says-openais-chief-research-officer/) with OpenAI’s Mark Chen defending the company’s disclosure practices around the recent agent-containment incidents, arguing the transparency is the point. That’s the morning. Argon’s selective release and the self-regulation pledge are two expressions of the same posture — frontier labs declaring themselves fit to decide who gets access to what, and when. Watch October 24 for Livenerf’s first real data point.