← Daily AI news

Daily AI news, Tuesday, 27 October 2026

Daily AI / Tech / VC Briefing: Tuesday 27 October 2026

OpenAI logo and brand image
Image: OpenAI (openai.com)

Good evening, here's today's briefing. OpenAI says its models have made genuine scientific discoveries and has launched a much cheaper workhorse model. Anthropic's IPO now has a target week and a clearer profit picture. Meta has lifted its share sale to $97.5 billion, with Nvidia putting in $3 billion, just as Microsoft, Alphabet and Amazon prepare to report. Cognition and SpaceX have made competitive moves in agents, a senior Meta AI leader has left to start a company, and a neocloud raised $100 million in a week when Musk predicted most of them will die.

LEAD: OpenAI says its models disproved the Exponential Time Hypothesis and proved a major maths conjecture, and launches GPT-6.1 Terra at half the price

Background: Claims of AI-made scientific breakthroughs have become a competitive front between the big labs. OpenAI used small groups of outside experts to guide and check its models' work. What happened: OpenAI announced three results. Its models proved Sogge's local smoothing conjecture, a central problem in harmonic analysis, in every spatial dimension (it had only been proved in three). They produced a claimed refutation of the Exponential Time Hypothesis, a widely relied-on belief about how hard certain computing problems are. And they produced rigorous results in quantum field theory. The maths and computer science proofs are public, including formal proofs written in Lean; the physics group has held its material back for now. Rochester mathematician Alex Iosevich called it a "historic development", while MIT's Ryan Williams said refuting ETH would be a major result and experts are working to verify it. Separately, OpenAI launched GPT-6.1 Terra, a cheaper everyday model now available in ChatGPT, Codex and the API. Its API price is $1 per million input tokens and $5 per million output, half that of GPT-5.6 Terra and GPT-6.1 Sol. OpenAI says it more than doubles its predecessor on the Terminal-Bench 4.0 coding test (50.3%) and costs about a ninth as much per task as Claude Opus 5.5 at maximum effort. It also rates the model "Critical" for cybersecurity capability under its own safety framework. Why it matters: If the ETH proof holds up, it is the strongest evidence yet that AI can produce original research, which strengthens OpenAI's pitch to investors. Terra matters more for buyers today: it continues the price war with Anthropic, and a cheap model that rivals OpenAI's own top tier on agent tasks puts pressure on every coding and agent product priced on frontier models. Sources: Al Jazeera · Engadget · OpenAI: GPT-6.1 Terra

Anthropic IPO update: listing targeted for the week of 16 November, up to $100 billion raise, and an adjusted profit

Background: Anthropic made its S-1 filing public on Monday, covered in yesterday's briefing. It plans to list on the Nasdaq as "ANT". What happened: The Wall Street Journal reports Anthropic is aiming to go public the week of 16 November, raising as much as $100 billion at a valuation of $1.5 trillion to $2 trillion. That could beat SpaceX's record $86 billion IPO in June. Fortune's read of the filing adds that most of the $80.7 billion first-half loss, $72.4 billion of it, was an accounting charge on convertible notes. The operating loss narrowed to $693 million, and on an adjusted basis Anthropic made $1.6 billion of operating income, about a 10% margin. In the second quarter it earned $11.6 billion in revenue against $5.6 billion of compute spending, or $2.08 of revenue per dollar of compute. Nearly 12,000 organisations now spend at least $100,000 a year with it. AFP notes the Pentagon has labelled it a security risk. Why it matters: An adjusted profit and a firm date make this a real test of public-market appetite for AI. AP notes it could become the biggest IPO ever, and some investors are already valuing it nearer $1.5 trillion than $2 trillion, Reuters reports. Sources: AP · Fortune · CNA/AFP · Reuters via Investing.com

Meta lifts its share sale to $97.5 billion with Nvidia putting in $3 billion, as Big Tech earnings arrive

Background: Yesterday's briefing covered Meta's plan to sell up to $95 billion of shares to fund AI data centres. Meta reported negative free cash flow in the third quarter for the first time in over a decade. What happened: Meta raised the total to $97.5 billion: $25 billion of common stock priced at $625 a share, about $12.5 billion of convertible preferred shares, and a $60 billion programme to sell shares over time from next year. Nvidia, AMD and Broadcom have shown interest in up to $5 billion of the stock, and Nvidia confirmed a $3 billion investment alongside a deal to work together on Meta's Muse assistant and models. Microsoft reports on Wednesday, for the first time with a new "Agents and Infra" division and actual Azure revenue figures. Alphabet and Amazon also report this week. Analysts expect Google Cloud revenue to grow about 95%, Azure about 48% and AWS about 38%. Why it matters: Chipmakers buying shares in their biggest customer adds to worries about circular money in AI. This week's cloud numbers will show whether AI demand is paying for the spending, and that feeds directly into how investors price the Anthropic listing. Sources: Reuters: Meta raise · Reuters: Nvidia and Meta · Reuters: earnings preview

Competitor moves: Cognition puts Claude Mythos inside Devin Security with "hundreds of millions" in free credits, and SpaceX pitches shared sales bots

Background: Cognition, maker of the Devin coding agent, passed $1 billion in annualised revenue in September and raised over $2 billion at a $48 billion valuation. Security has become a major selling point as frontier cyber models get more capable. What happened: Cognition upgraded Devin Security so it runs deep vulnerability scans using Anthropic's Claude Mythos and other frontier cyber models. Enterprise customers get those models through Devin without needing their own access. Devin keeps a living threat model of the codebase, re-scores findings by how exploitable they really are, and can run thousands of parallel virtual machines to fix them. Cognition is also committing hundreds of millions of dollars in Devin credits for free scans and remediation. Separately, SpaceX's AI unit (SpaceXSI) pitched shared Grok Bots for sales teams, with ready-made bots for account research, quoting, deal review, competitive intelligence and sales training. In smaller news, Cursor added DeepSeek V4.1 Flash as a cheap option it says costs $0.60 to $0.98 per task on its own benchmark. Why it matters: Cognition is turning access to restricted cyber models into a product advantage that Cursor, GitHub Copilot and Claude Code don't yet package in the same way. Free credits on that scale are a land grab for enterprise security budgets. SpaceX's post is a public pitch for team agents aimed at the same business customers as Anthropic and OpenAI. Sources: Cognition · SpaceXSI · Agentdex roundup

Key hire: Meta Superintelligence Labs director Rohit Patel leaves to take on Office and Workspace

Background: Meta has lost several senior people recently, including Joel Kaplan, Andrew Tulloch and Emily Dalton Smith. What happened: TechCrunch reports Rohit Patel, a director at Meta Superintelligence Labs who spent eight years on AI at Meta, has left to launch Ndoc. It is pitched as an AI-native replacement for Microsoft Office and Google Workspace, with presentation, narrative and model views in one document and an open-source file format. It is built on open-source models, opens an enterprise pilot and alpha waitlist today, and has a team of three with no outside funding yet. Why it matters: It's another senior exit from Meta's AI group during a costly push, and another bet that AI work tools need to be built from scratch rather than added to existing office apps. Sources: TechCrunch

VC: General Compute raises $100 million at $800 million, days after Musk predicted mass neocloud deaths

Background: CoreWeave's shares are at a 12-month low, and Elon Musk has just warned that most of the 200-plus "neoclouds" (AI chip rental companies) will die. What happened: General Compute, which builds inference capacity from non-Nvidia chips (SambaNova, Positron and Cerebras) partly in converted bitcoin-mining sites, raised a $100 million Series A led by Blackbird Ventures and Abstract Ventures. It is valued at $800 million, up from $60 million after its May seed round, on the back of a customer deal worth $100 million in annual recurring revenue. In India, Nvidia backed Simplismart, whose software runs AI models on fewer GPUs, in a $14 million Series B at about $100 million. Why it matters: Investors are still funding compute, but they are shifting from GPU landlords to cheaper inference and efficiency plays, the part of the market expected to survive a shake-out. Sources: TechCrunch: General Compute · TechCrunch: Simplismart

← Monday, 26 October 2026 Wednesday, 28 October 2026 →

Get the weekly sift

The best new AI and agent releases, with one-line verdicts. Occasional picks, no spam.

Privacy policy: your email goes to our email provider, Resend. Unsubscribe any time.

News is editorial and never paid for. Sponsored content on ToolSift is always labelled.