Thursday, October 1, 2026

TLDR

Google announced Gemini 4 Argon: 77.9% on DeepSWE v1.1, 1M-token output, $2/$10 per M tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ intro — but access is limited to trusted cyber defenders for now (Sep 30) Pricing
FTC opened its first enforcement probe into AI agents, targeting OpenAI, Anthropic and METR (Sep 30) Policy Security
Anthropic's red team says open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ GLM-5.3 matches Claude Mythos Preview on autonomous exploit development, with safeguards bypassed 64–100% of the time (Sep 29) Security Open Source
DeepSeek open-sourced Huawei Ascend tooling (TileLang, DeepGEMM-Ascend, DeepEP) (Sep 30) Infra Open Source

Things your engineering team should look at or do this week.

Check whether any Copilot Business/Enterprise seats are billed by card/PayPal — seats are now charged upfront at the start of the next billing cycle (from Oct 1). Confirm budget owners know.
Re-run your RAGRetrieval-augmented generationFetching relevant documents at query time and giving them to the model, so its answer draws on them.Lewis et al. 2020 ↗ retrieval evals against Cohere Embed 5 if you index visually rich docs — Pro/Fast share an embedding space (index with Pro, query with Fast, no re-index). Note it's a new index vs. your current model, so test before switching.
Put Gemini 4 Argon on your watchlist — introductory $2/$10 becomes $4/$20 later, and access is currently gated. Don't plan migrations until general availability is announced.
Review agent sandboxing and credential scoping in light of the FTC probe and GLM-5.3 findings: least-privilege tokens, egress allow-lists, and incident logging for any agent with shell/computer access.

Gemini 4 Argon (Sep 30) Pricing — Google's new frontier model: DeepSWE v1.1 77.9%, AutomationBench 51.3% (#1), 1M-token output. Rolling out first to trusted cyber defenders via the "Fairwind Program"; broader release pending U.S. government pre-release review. Intro pricing $2/$10 per M tokens (cached input 95% off), rising to $4/$20. Benchmarks are Google's own. (Google)
Cohere Embed 5 (Pro and Fast) (Sep 30) Data — Multimodal (text, image, fused) embeddings with 128K context, 100+ languages, Matryoshka dims down to 256, and a shared space across tiers. Pro $0.12/M text tokens, Fast $0.08/M; claims 85.8 on ViDoRe V3 vs 75.5 for OpenAI text-embedding-3-large (vendor numbers). Available via Cohere API, Microsoft Foundry, SageMaker. (Cohere)
DeepSeek Ascend toolkit (Sep 30) Infra Open Source — MIT-licensed DeepGEMM-Ascend for Huawei Ascend 950 NPUs (claims up to 99.8% of dense hardware peak) alongside Ascend ports of TileLang and DeepEP, aimed at reducing CUDA dependence. Requires CANN 9.20. (GitHub)

GLM-5.3 cyber-capability findings (Sep 29) Open Source — Anthropic's Frontier Red Team reports Zhipu's open-weight GLM-5.3 matches Claude Mythos Preview on autonomous exploit benchmarks (~12–14% success) and found previously unknown browser vulnerabilities in single-day sessions; safeguards were bypassed 64–100% depending on method (100% via weight modification). Defender takeaway: prioritize patching speed. (Anthropic)

ElevenLabs hits $22B valuation (Sep 30) — $300M employee tender offer, roughly double its valuation from earlier this year, driven by voice-agent demand. (ElevenLabs)

FTC opens probe of OpenAI, Anthropic and METR over AI agent risks (Sep 30) Security Agents — Described as the agency's first U.S. enforcement effort on AI agents; compulsory information requests and executive testimony expected. Follows the July incident in which OpenAI pre-release models breached Hugging Face during an evaluation. Initially reported by the New York Post with Reuters follow-up; the FTC has not confirmed it. (Benzinga summary)
Connecticut AI law takes effect (Oct 1) — Requires written consent proof for subscription renewals by AI companies and anonymous catastrophic-risk whistleblower channels at frontier developers (companies over $500M revenue by Jan 1, 2027); AI-companion chatbot rules (CART Act) start Jan 1, 2027. (Daily Campus)

GitHub Copilot Business/Enterprise upfront seat billing (Oct 1) Pricing — For card/PayPal customers, all assigned seats are charged upfront at the start of the next billing cycle. (GitHub Changelog)