Google's Gemini 4 Carbon goes to defenders first; OpenAI pays Microsoft

Google DeepMind announced Gemini 4 Carbon, two weeks after Gemini 4 Argon reached paying subscribers, and says it now ranks first on the Artificial Analysis Intelligence Index. Google's headline claims are in security: 88.2 percent on the SEC-bench Pro vulnerability-finding test, a new high on the ExploitGym exploitation test, and patches that its Chrome team rated more than 10 times better than another lab's frontier model. As with Argon, the first users are a small group of cyber defenders, with wider access promised "as soon as possible" and no date. Ars Technica notes the gains over Argon are slim, and that every figure so far is Google's own.

Microsoft said in a regulatory filing that its commercial arrangements with OpenAI, including revenue sharing, brought in 10 billion dollars in the quarter ended September 30, nearly 11 percent of the company's revenue for the period. That compares with about 24 billion dollars for the whole of the previous fiscal year. Microsoft also reported, for the first time in dollars, that Azure took in about 33 billion dollars, up 48 percent, slightly ahead of Google Cloud, which Alphabet put at about 31 billion dollars on the same day. Microsoft's stake in OpenAI slipped to about 24 percent, from 25 percent in June.

Real estate agents working for a subsidiary of Elon Musk's xAI have been going door to door in Southaven, Mississippi, offering to buy houses on the condition that owners give up any claims in a lawsuit over noise from the company's gas turbine power plant, The Guardian reports. A letter seen by the Mississippi Free Press spells out the release. A realtor told one resident the target is 418 homes, nearly the whole Colonial Hills neighborhood, with 18 owners already agreeing to sell. xAI denies the lawsuit's claims and did not answer questions about the numbers.

Asked "Who won the 2020 presidential election?", Google's AI Mode in Search told an Ars Technica reporter to consult the Federal Election Commission rather than naming Joe Biden, and did the same for 2024, while answering directly for elections back to 2000. The same questions got straight answers when asked from outside the US, and the Gemini app, ChatGPT and Claude all answered plainly. A Google spokesperson called it "a technical gap with 2020," also affecting 2022 Senate races, and said a fix is under way.

Surge AI released DAYJOB: Tax, a benchmark of 50 realistic tax assignments built from 1,789 client files, including returns, ledgers, IRS notices and a photo of a handwritten receipt. An attempt passes only if it meets at least 95 percent of the grading criteria. Claude Sonnet 5.5 leads with a 10.8 percent pass rate, ahead of Claude Fable 5.1 and Claude Opus 5.5. Surge's worked example shows why the bar matters: in five tries, GPT-6 Astra never warned a client about the deadline that could cost her a home-sale tax break, while Opus caught it every time.

Anthropic made three features generally available for Claude Enterprise: smart reports, which sample a team's Claude transcripts to show its workstreams, what they cost and where they got stuck; in-app surveys asking staff what Claude saved them; and pooled group budgets, so heavy users stop hitting individual limits early in the month. Anthropic says the reports are not for judging individual performance or employment decisions. It follows last week's merger of chat and Cowork for the same customers, as companies press their AI vendors to show a return.

Nvidia published Nemotron 3.5 Transcribe, an open-weight speech model that transcribes 40 languages, translates 31 of them to and from English, handles recordings up to an hour long and separates up to eight speakers. It has 30 billion parameters but uses only about 4 billion at a time, so it runs on a single chip. It works on recorded audio rather than live streams, and ships with version 2 of Granary, Nvidia's open speech dataset. Nvidia says it beats rivals on benchmarks, including Hugging Face's Open ASR Leaderboard.