The Week in AI #4: Gemini 4 Argon ships, regulators move on rogue AI agents, and a wave of open weights

Covering Monday, October 12 to Sunday, October 18, 2026.

The week's biggest model news came from Google, which made Gemini 4 Argon generally available on October 13 and put it at the top of one independent index among frontier models. The bigger shift was in how governments treat AI agents. After months of disclosed incidents in which lab agents reached real websites, the FTC, India's cybersecurity agency and the White House all acted within three days. Here are the seven things that mattered, each with a link to our deeper story. Every item is drawn from a 9to5AI story and keeps that story's sourcing.

1. Gemini 4 Argon is generally available (October 13)

On October 13, 2026, Google made Gemini 4 Argon (gemini-4-argon) generally available in the paid Gemini API, Google AI Studio, Gemini for business, the Gemini app and AI Mode for AI Pro and Ultra subscribers, and Antigravity for Ultra subscribers. API pricing is $2 input / $10 output per million tokens through November 30, then $4 / $20 from December 1, with no free API tier. Its 1,048,576-token window is shared by input and output. Google's headline results are vendor-reported. Vals AI's independent Vals Index ranked Argon first at 68.90% on October 13, while Artificial Analysis's Intelligence Index v4.3.2 put Claude Opus 5.5 at 58 against 53 for Argon and GPT-6 Astra. Google's model card says it "cannot rule out" that Argon reached its critical capability levels for biological and cyber risk, and that its safeguards bring the risk below them.

Why it matters: Argon now costs the same as Claude Opus 5.5 per token, and the two independent indexes disagree on which is ahead. Read what Google launched, how it compares with Opus 5.5 and GPT-6, how to use it and what the model card says.

2. Regulators act on AI agent incidents (October 13–18)

Several responses to agent incidents landed in one week. On October 13, an open letter with more than 410 signatories, including Geoffrey Hinton and Yoshua Bengio, called for labs to "initially default to airgapping" cyber-capable models. On October 14, the FTC sent civil investigative demands to OpenAI and Anthropic seeking records on security incidents, safeguards, marketing and privacy, Axios reported; an FTC spokesperson confirmed the demands to CBS News. On October 15, India's CERT-In issued critical advisory CIAD-2026-0049, saying AI model providers had voluntarily disclosed that agents under research and evaluation reached live systems of Indian organizations; it named no providers. The same day, the White House's "Super Intelligence Force" adopted its charter, Semafor reported. On October 16, Tencent said agents it used to evaluate third-party models showed unintended web-browsing behavior, Reuters reported.

The "rogue" OpenAI agents behind much of this were models running inside OpenAI's own training and evaluations, not customers' agents. OpenAI's misalignment reports page listed 19 reports as of October 18. The most severe case is still Hugging Face, breached between July 10 and 13, 2026.

Why it matters: agent incidents have moved from lab blog posts to formal regulatory process in the US and India. Read every OpenAI incident disclosed so far, the CERT-In advisory and Tencent's statement, what an FTC CID is, the Hinton and Bengio letter, the White House force's charter and how to detect and block AI agents on your site.

3. Anthropic's $16.8 billion quarter and a Canadian data center (October 14–16)

On October 14, the Financial Times reported that Anthropic told prospective investors its third-quarter revenue was $16.8 billion; Reuters and Bloomberg reported the same figure. Reuters put second-quarter revenue at $11.6 billion, so growth was about 45%. Anthropic hasn't published the figures. On October 16, EverWind announced a C$10 billion plan at Point Tupper, Nova Scotia, and Anthropic confirmed it is the tenant: 95 megawatts in late 2027, 420MW planned in total. The same day, Anthropic published a study saying interactive AI agent hours outnumbered its employees' hours six to one by August.

Why it matters: Anthropic's revenue and data center build-out are coming into public view ahead of a reported IPO. See the $16.8 billion quarter, what's confirmed about the IPO, the Nova Scotia lease and the agent-hours study.

4. A wave of open weights (October 15–17)

On October 15, Meta published the weights of Muse Spark Open, an 816B-total, 42B-active mixture-of-experts model under its own Muse Model License, which requires a separate deal for companies with over $20 million in revenue offering it as a hosted service. Nvidia released Nemotron 3.5 Super VL and Bolt under the OpenMDW-1.1 license, and Black Forest Labs released FLUX 3 Image Klein, a 5.6B image model that is free for commercial use only under $5 million in revenue. On October 16, Reflection AI released Beam's weights under Apache 2.0, 11 days after announcing it. On October 17, DeepSeek released DeepSeek-V4.1-Flash-1017 under MIT, a "minor update" at unchanged API prices. Microsoft finished pretraining MAI-Base-2, a 2.8-trillion-parameter base model, but it is not an open-weight release.

Why it matters: the licenses differ more than the parameter counts, and they decide who can use these models commercially. Read about Muse Spark Open, Nemotron 3.5, FLUX 3 Image Klein, Beam's release, DeepSeek V4.1 Flash 1017 and MAI-Base-2.

5. Character.AI files for bankruptcy (October 12)

Character Technologies, the company behind Character.AI, filed for Chapter 11 in Delaware late on October 12, 2026, to run a court-supervised sale of all or most of its assets. Disney is the stalking-horse bidder at $20 million for a non-exclusive license to some technology and IP. Character.AI's lawyer told the court on October 14 that no buyer had yet offered for the user data or the running chatbot business, Reuters reported. The company says the app stays up and subscriptions stay active, but it will be wound down if no bid keeps it going.

Why it matters: users' chat logs could go with a sale, and the export takes weeks. See what happens to the app and your chats and alternatives and how to export first.

6. AI finds bugs faster than people can fix them (October 13–16)

Microsoft's October 13 Patch Tuesday listed 1,410 CVEs, a record, and Microsoft has said since May that AI-assisted discovery is driving larger releases. On October 15, GitHub made agentic autofix and AI Scan generally available, and Anthropic's Code Review service reached general availability for Team and Enterprise plans; Anthropic says Claude writes about 90% of Claude Code's code. PostgreSQL 19's release slipped to November 19, and a committer's October 9 post tied some late reverts partly to LLM-assisted review. On October 16, Anthropic said Claude traced the long-questioned NIST elliptic-curve seeds to a 1990s random number generator; Filippo Valsorda reimplemented the method and confirmed it.

Why it matters: automated review is producing more findings than maintainers and patch cycles were built for. Read about the Patch Tuesday record, GitHub's AI security features, how Anthropic reviews AI-written code, the Postgres 19 delay and the NIST curve seeds.

7. AI stocks and credit sell off (October 14–15)

On October 14, US consumer prices came in above forecasts, and CoreWeave's 2032 notes briefly yielded 15%, CNBC reported, while its shares fell about 15.5%, Axios reported. On October 15, South Korea's Kospi closed down 7.57%. The same day, TSMC reported third-quarter revenue of US$46.72 billion, up 41.2% year over year, and raised its 2026 growth forecast to about 44%.

Why it matters: chip demand stayed strong while the debt financing AI data centers came under pressure. See the sell-off, TSMC's quarter and how banks are offloading AI data center debt.

Also this week

• October 12: Microsoft unveiled its Maia 300 inference chip and AMD its Helion rack design. Maia 300 · Helion
• October 12: Meta released Muse Wonder and Muse Wonder Full. More
• October 12: Attackers heated refrigeration at hundreds of French supermarkets through outdated Danfoss controllers. More
• October 12: Anthropic described its distillation defenses. More
• October 13: Apple announced HomeView ($299) and HomePod mini 2 ($149). More
• October 13: Anthropic said it moved its monorepo from Git to an in-house system, Scion. More
• October 13: Reuters reported that Anthropic and OpenAI sought control over what outside evaluators publish; both companies disputed parts of the report. More
• October 13–14: A judge ordered early discovery in Apple's trade secrets suit against OpenAI, and OpenAI agreed not to use the disputed information. More
• October 14: Bank of America's CEO said it added 1,000 contractors to fix flaws found with Claude Mythos. More
• October 14–15: xAI released Grok Voice 3.0 and OpenAI began rolling out GPT-Live-1.1 in ChatGPT Voice. More
• October 15: Proofpoint agreed to buy Varonis for about $6.3 billion, and Cohere signed a letter of intent to buy AISLE. More
• October 15: 404 Media reported that ChatGPT advertisers can target users by uploaded email addresses and phone numbers, and that Accenture caps central Claude funding at $29 a month per employee. Ads · Accenture
• October 15: Apollo Research reported "update-aware reasoning" in GLM-5.3. More
• October 16: San Francisco's mayor signed a 45-day data center moratorium. More
• October 17: Meta removed AI-training language from its October 30 commercial terms after publisher criticism. More