Two weeks after Argon, Google has announced Gemini 4 Carbon. It takes first place on the Artificial Analysis Intelligence Index v5 with 59 points, against Argon's 50 (The Decoder). Like Argon, it rolls out in stages, and almost nobody can use it yet (Ars Technica).

Sold on cyber defense

• Finding bugs: 88.2 percent on SEC-bench Pro, first among tested models. 54.8 percent on ExploitGym, which tests turning a bug into a working exploit. Google ran several security tests without safety guardrails.
• Secure code: Google says it is twice as likely as Argon to write vulnerability-free code, and the Chrome team claims patches 10 times better than non-Google models.
• Long jobs: Google's teams ran it for days on root-cause analysis and bug fixes in large codebases.
• Prompt injection: it blocked every attack in Gray Swan's test (Argon 0.7 percent, Opus 5.5 1 percent, GPT-6 Astra 8.5 percent got through). The test looks saturated; the problem is not solved.

The fine print

• The lead over GPT-6 Astra and Claude Opus 5.5 is a few points, and Ars Technica says Carbon could be a post-training upgrade of Argon.
• Hallucination rate rose from 4 to 11 percent, though accuracy rose from 48 to 58 percent.
• In agentic coding it still trails Claude Code with Opus 5.5 by three points.
• It thinks longer: 2.3 times the reasoning tokens of Argon, and over 50 minutes per legal research task on Vals AI.

Who gets it, and at what price

First, cyber defense partners in Google's Fairwind program (SiliconANGLE), after the US government's voluntary testing framework. No date for a wide release. The introductory API price matches Argon, $2 per million input tokens and $10 per million output, and will double later; Google declined to say when.

Microsoft shows Azure in dollars

• Azure brought in $33.2 billion in July to September, up 48 percent, the first time Microsoft has given a dollar figure. Google Cloud brought in $30.7 billion and grew about twice as fast (TechCrunch).
• Microsoft recorded $10 billion of revenue from OpenAI in the quarter, nearly 11 percent of the total; its stake slipped from about 25 to 24 percent (TechCrunch).

Claude Code's hidden token bill

A write-up that topped Hacker News says that every ~30 seconds, Claude Code sends each background agent's whole transcript to that agent's model, to produce a 3-5 word status line. In the authors' telemetry it was 16.4 percent of Claude Code spend, more than the main conversation (15.7 percent). A GitHub issue lists four more measurements of 18-26 percent. There is no setting to turn it off. These are users' measurements, not an official number, and the authors sell an observability tool.

AI agents hacked Anthropic

Security startup Cathedral says its AI agents reached Anthropic's internal systems through third-party sales and admin software in under twelve hours. Anthropic confirms access to a limited set of sales, HR and IT systems, says chats, source code and model weights were not exposed, and says Cathedral broke its bug bounty rules (The Decoder, citing The Washington Post). Separately, Anthropic's S-1 warns that the gas buildout for data centers could raise prices and squeeze its fuel supply (Wired).

Google opens the doors

• Android Studio now supports more than 40 outside coding agents over the Agent Client Protocol; a new `android preview` command made agents 77 percent faster on UI tasks, by Google's unexplained measure (The New Stack).
• Google's Connected Apps Platform lets developers plug apps into Gemini and Search through MCP, in a US preview (TechCrunch).

Builder's bench

• GitHub Copilot Fusion picks and chains models for you; GitHub says it matched Claude Code with Opus 5.5 Max at about half the API cost (The New Stack).
• GitHub load: 7.4 billion commits a month, up from 2.9 billion in August, mostly from agents; 4 million CPU cores added this year (The New Stack).
• Nemotron 3.5 Transcribe: Nvidia's open speech model, 40 languages, translation, speaker labels, 30B parameters with about 4B active, one GPU (The Decoder).
• Artificial Analysis v5: 60 percent private tests; the open-model gap grew from 3 to 12 points (The Decoder).

Security

• PHP2Shell: two flaws rated 9.2, confirmed exploited by CISA; Censys counts 32.51 million PHP services on affected versions (The Hacker News).
• Volt Typhoon ran the Silent Train relay network of hacked SOHO and IoT devices for four zero-day campaigns (The Hacker News).
• Microsoft Authenticator for Android: CVE-2026-80097 let a no-permission app pose as Teams and take work tokens; fixed in 6.2606.4246 (The Hacker News).
• AI-run attacks: a CyberStrikeAI setup split work across 44 agents against targets in Japan (The Hacker News); in South Korea, Daishin, Mirae and Bithumb were hit, with 38,490 Daishin customers exposed (TechCrunch).

One thing to do today: if you run PHP, upgrade to 8.5.12, 8.4.27, 8.3.36 or 8.2.35.

Sources covered

1 story · 96 sources (66 news, 25 company, 5 voices)

1. Claude Code regularly sends each background agent’s transcript to its model
   📰Hacker News↗🎙️Simon Willison↗🏢LangChain↗📰RIA Novosti↗📰Facebook↗📰MotorBiscuit↗📰TechCrunch↗📰The Hackers News↗📰TechCrunch↗📰CNBC↗📰dbt Labs Blog↗📰CNBC↗📰The Hackers News↗📰The Verge↗🏢Vercel AI↗🏢AWS AI/ML↗🏢AWS AI/ML↗🏢AWS AI/ML↗🏢Vercel AI↗🏢Databricks↗📰TechCrunch↗📰TechCrunch↗📰TechCrunch↗📰The Verge↗🏢LangChain↗📰TechCrunch↗🎙️Simon Willison↗📰MIT Technology Review↗📰The Verge↗🏢Meta AI↗🏢GlobalFoundries↗📰SiliconANGLE↗📰TechCrunch↗📰The Decoder↗🏢NVIDIA↗🏢Hugging Face↗📰Ars Technica↗📰The Verge↗🏢ClickHouse↗🎙️Simon Willison↗📰Ars Technica↗📰The New Stack↗🏢OpenAI↗🎙️Ben Lorica↗📰Hacker News↗📰Google Chrome Blog↗📰The Verge↗📰TechCrunch↗📰The Decoder↗📰BleepingComputer↗📰The Decoder↗📰The Decoder↗📰The New Stack↗📰The Verge↗📰The New Stack↗📰TechCrunch↗📰The Decoder↗📰TechCrunch↗📰The Decoder↗📰SiliconANGLE↗📰TechCrunch↗📰The New Stack↗📰The Verge↗📰LWN.net↗🎙️Simon Willison↗🏢OpenAI↗📰InfoQ↗📰BleepingComputer↗🏢Vercel AI↗🏢OpenAI↗📰SiliconANGLE↗📰SiliconANGLE↗📰TechCrunch↗📰The New Stack↗📰SiliconANGLE↗📰MIT Technology Review↗📰The Verge↗📰The New Stack↗📰The Verge↗🏢Vercel AI↗📰The Decoder↗🏢Supabase↗🏢Supabase↗🏢Vercel AI↗📰The Decoder↗🏢ClickHouse↗📰Wired↗📰dbt Labs Blog↗📰dbt Labs Blog↗📰Wired science↗📰The Hackers News↗📰dbt Labs Blog↗📰dbt Labs Blog↗🏢Temporal↗🏢Elastic↗🏢MotherDuck↗