Daily AI digest, November 6

Felix Laurent 6 Nov 2026 4 min read

News

The biggest AI news of November 6 was about who still wants to pay for AI compute. Microsoft is winding down its deal to rent AI capacity from Oracle. Investors told The Wall Street Journal that Anthropic's IPO valuation will likely have to come down. A startup said a non-public Anthropic model proved a math conjecture that had been open since 2008. Below are the big stories of Friday, November 6, plus the late Thursday news and the day's rumours.

Microsoft winds down its Oracle AI cloud deal

Microsoft declined to renew its contract to rent AI computing capacity from Oracle, Bloomberg reports. It has been switching off the rented servers in recent months. Oracle shares fell 3% to 4% on Friday, while Microsoft barely moved.

The report came a day after Bloomberg said Microsoft is also stepping back from deals with smaller GPU clouds. Our breakdown of what Microsoft is cutting is here.

Anthropic's IPO valuation comes under pressure

Three large investors who planned to buy into Anthropic's IPO told The Wall Street Journal the target will likely come down after the tech selloff. Anthropic had hoped for about $1.5 trillion, and one investor put the value "closer to $1 trillion". Anthropic declined to comment.

The IPO is still planned for this month. We covered the investors' comments here, and the $895B of compute commitments behind them in Thursday's article.

OpenAI tests agents on everyday lab work

OpenAI published BioBasicsBench, a benchmark of whether AI agents can read raw results from routine wet-lab work, such as gel images, colony plates and microscopy. GPT-6 Astra earned 66.7% of the available credit at its highest common reasoning setting. GPT-5.6 Sol, the previous generation, earned 43.8%.

OpenAI built the benchmark with Ligo, a startup that designs enzymes with AI. More in our article.

A startup says a non-public Anthropic model proved a 2008 conjecture

PSI, a research startup, says its internal system found a proof of the Aaronson-Ambainis conjecture about quantum speedups on random inputs. The system ran instances of a non-public Anthropic model, and the Lean proof assistant accepted the final proof. PSI says physicist Jim Halverson audited it.

This is the company's own report, and the proof has not been peer reviewed. Stanford researchers who settled a special case in August have read it and are writing an exposition.

Smaller updates

Three smaller stories came out on Friday.

US prosecutors charged three foreign nationals with using AI to impersonate executives and attorneys and steal more than $129M. The scheme ran from July 2024 to February 2026 and hit banks and hospital systems, NBC News reports.
Zyphra released MAIA in beta, a team AI agent that works in Slack and a Mac app and remembers what a team decides. More in our article.
Swiss startup Embodied AI unveiled Air, a wheeled humanoid that learns factory tasks from human demonstrations. The company says it hit 98% reliability on a cable-handling test, The AI Insider reports.

Two of the three are about AI that works alongside people at their jobs.

Late Thursday updates

Nine stories landed after our Thursday digest went out.

Google released Gemini Omni 1.2 Flash, a video model that makes 15-second clips at $0.07 per second at 720p, 30% cheaper than before. We covered it here and compared it with Seedance and Wan here.
Thinking Machines Lab released Inkling 2 Small, a 594B model with 17B active parameters that scores 31 on the Artificial Analysis Intelligence Index. Apache 2.0 weights are promised for next week; more in our article.
Cursor launched Apps, which lets cloud agents build and host live apps with a database and sign-in. Details are here.
Meta paused indefinitely its new terms that would have let it scrape the websites of business account users. Courts in Norway and Brazil had issued injunctions, and the UK CMA had opened a probe, which we reported on October 30.
The Justice Department launched a Super Intelligence Task Force to pursue foreign actors and fraudsters who misuse AI.
SoundHound AI is cutting about 27% of its workforce to save $120M a year after its LivePerson acquisition, Crypto Briefing reports.
A researcher used AI agents running about 17,600 login tests to find an OpenSSH flaw on Alpine and other musl-based Linux systems. It is fixed in OpenSSH 10.6p2, BleepingComputer reports.
Check Point published a ChatGPT Lockdown Mode bypass that could leak full chat histories. OpenAI had already fixed it on September 18, GBHackers reports.
We published a deeper read of METR's assessment that AI is roughly doubling the pace of research at Anthropic. It is here.

Four of the nine are new models and agent tools, released the evening after the selloff.

Rumours

Four more stories circulated on November 6, but we could not read the primary report or confirm them with the companies.

Nvidia is reportedly requiring Southeast Asian cloud partners to enroll in Fleet Intelligence, which logs where its chips sit, Bloomberg reports. The aim is to stop smuggling to China.
The incoming Democratic Congress reportedly plans closer oversight of Nvidia's China exports and circular AI financing, Bloomberg reports.
TSMC is reportedly weighing 3D DRAM manufacturing at Nvidia's request amid the memory shortage, The Information reports.
AI training bots are reportedly overwhelming university libraries and museums, with traffic surges above 1,000%, CBC reports.

Treat all four as unconfirmed until the companies announce them.

What the day was about

Friday showed the buyers of AI compute getting pickier. Microsoft stopped renting Oracle's servers, and investors want Anthropic's IPO priced closer to $1 trillion than $1.5 trillion.

The research did not slow down. GPT-6 Astra scored about 23 points higher than its predecessor on lab results. A startup says a non-public Anthropic model proved a conjecture that had stood since 2008. On November 6 the money got more careful while the models kept improving.