The day's tech, sifted: Nov 10, 2026

What matters today: Anthropic launched its IPO roadshow at $70 to $78 a share, a valuation of up to $1.4 trillion, one day after shipping Claude Fable 5.2 and Mythos 5.2 at 20% lower prices. Google answered by opening Gemini 4 Carbon to paying developers, and DeepMind says its agents verified Signal's core encryption in Lean. In security, the OpenSSH-on-musl flaw *sshole is under active attack.

AI / LLMs

Gemini 4 Carbon goes wide: Google's top model, limited to cyber defenders for two weeks, is now open to paid API and AI Ultra users, with a preview "swarm" feature in Antigravity for running groups of agents. Google positions it above last month's Argon, which stays the default for most coding and knowledge work (The Verge); it has a 1M-token context window and is half price on Vercel's AI Gateway through Nov 30. Why it matters: all three frontier labs now sell a security-capable top tier, and Google's reached general developers a day after Anthropic's.
Claude Fable 5.2 and Mythos 5.2: Anthropic's flagship pair launched Nov 9 at $8/1M input and $40/1M output tokens (20% below 5.1; fast mode is $16/$80). Fable is generally available; Mythos is limited to vetted cyber and life-sciences organizations. Anthropic says about 1,000 Fable agents formalized Fermat's Last Theorem in Lean in roughly 30 hours. Outside reads so far: Artificial Analysis scores it 65, its highest yet, and Scale Labs and CAIS say it completes 41.3% of real freelance projects at professional quality, about double Opus 5.5 (21.3%) and GPT-6 Astra (20.8%). Why it matters: the Fermat result is a vendor claim, but a price cut plus a doubling on a third-party work benchmark is a real shift in what agents can be handed.
Anthropic's office tools: Claude gets Sheets, Whiteboards and Claude Data (dashboards built from sources like Snowflake and Salesforce), plus connectors that embed partner apps such as Adobe and Atlassian inside Claude; a meeting assistant is "coming soon". Why it matters: Anthropic is moving from the coding niche into the everyday office software Microsoft owns.
Meta Muse Ember: Meta Superintelligence Lab's smallest model, claimed to beat Haiku 5.5 and GPT-6 Luna on intelligence per dollar, aimed at cutting agent costs. Why it matters: the cheap tier is where agent workloads actually run, and Meta is now competing there directly.
Gemini Robotics 2.5: DeepMind's new robot-control model was tested on 200+ tasks (tongs, pipettes, a humanoid tying a bow) and did real work in Mercedes-Benz factories; it ships with ASIMOV-jailbreak, an open safety benchmark. Why it matters: factory pilots and a public safety test are the two things robot models have lacked.
Agents meet a wall: ChatGPT dots can now link several local machines at once, while eBay appears to have blocked third-party agents, Meta's Muse included, from browsing listings. OpenAI also described Control Room, an internal system that can halt model workloads quickly on safety concerns. Why it matters: agents are gaining reach faster than sites and labs agree on where they may go.

Security & Privacy

*sshole exploited: attackers are abusing a critical, decade-old flaw in the OpenSSH and musl combination that gives unauthenticated access to privileged accounts on affected Linux systems. GreyNoise says a group it tracks as GT10 ran "mature intrusions" against its honeypots four days after disclosure. Why it matters: the patch window closed fast; check musl-based hosts (Alpine-style images, embedded devices) first.
Fireworks breach reaches Notion: this morning the AI infrastructure provider said hackers accessed some customers' datasets, incomplete model weights and billing data over the past month; by evening Notion had posted its own notice of a security incident involving a third-party AI vendor. Why it matters: a breach at one model host now travels downstream to the products built on it.
Claude Code settings leak secrets on GitHub: Apiiro found thousands of database passwords, API tokens and root logins in committed Claude Code settings files, because "don't ask again" stores the approved shell command verbatim. Why it matters: a token typed into an auto-approved command can leave with the repo.

Devtools & Infra

Plumbing for agents and GPUs: Nvidia detailed Agent Substrate, which adapts Kubernetes to agents that sit idle between turns, and open-sourced Snapshot, which restores initialized GPU workloads in seconds instead of minutes. Meta detailed TLX, a Triton extension giving engineers low-level control of GPU execution on Nvidia and AMD chips. Why it matters: cold starts and idle sessions, not model quality, are what make agent fleets expensive to run.
Compute bets, and one writeoff: SpaceX plans a gas-fired power plant and compute facility at Starbase (about $24B in initial phases), while SoftBank wrote off $3.7B of non-refundable advances for data center equipment after reviewing its plans. Why it matters: the buildout continues, but the first large prepayments are being abandoned.

Startups & Industry

Anthropic's IPO gets a price: yesterday's filing drew a correction; today the roadshow is official. Anthropic and holders will sell 825M Class A shares at $70 to $78 each on Nasdaq under "ANT" (roughly $58B to $64B at that range), implying a valuation of up to $1.4 trillion, below the $2 trillion once expected. The seven co-founders will sell $1.75B of stock between them, TSMC plans to buy up to $500M at the IPO price (and as much again in OpenAI), and Goldman reportedly projects revenue of about $52B this year. Why it matters: one of the largest listings ever now has terms, and they came in under the hype.
Nvidia, the delta: Trump says the FBI is "looking into" Huang's affair, a shareholder lawsuit accuses Nvidia of concealing it, and Ars notes the company has no obvious successor. On export controls, Bloomberg says Nvidia's 2025 figures undercounted chips sold to Megaspeed by 50K and that Moonshot rented China-based servers with B300 chips, possibly breaching US rules. Why it matters: the story has widened from a personal scandal to succession, litigation and chip-export reporting.
Mercor doubles to $20B: the AI training startup raised a $750M Series D from Benchmark, General Catalyst, Nvidia and others, up from $10B in October 2025. Why it matters: paying human experts to teach models is still one of AI's fastest-growing businesses.
States move on AI oversight: New York is launching a center to independently test AI systems and train safety staff, billed as a US first, and Australia is drafting an AI safety regulator with penalty powers after an OpenAI model hacked a government site in September. Why it matters: testing is shifting from lab self-reports to public bodies.
Also: IBM is ending H-1B and green card sponsorship for most sponsored roles in its consulting arm, and Apple now sells a $49.99 Apple Intelligence usage pack in Logic Pro. Why it matters: both are costs moving onto workers and users.

Research

AlphaProof Code: DeepMind says its agents wrote three million lines of Lean proofs to verify Signal's core encryption. Why it matters: machine-checked proofs of real, widely deployed software have been too labor-intensive to attempt at this size; if this holds up, verification becomes something you can buy with compute.
Epoch's Inference Performance Index: Epoch scores chips on memory and interconnect, not just raw compute, and says GB300 serves 25x an H100's load. Why it matters: raw compute comparisons understate how much newer racks help for serving.

Elsewhere

Courts and surveillance: EFF says the Supreme Court's Cisco v. Doe ruling effectively ends Alien Tort Statute suits against tech firms accused of aiding abuses, and EFF and the ACLU asked the Sixth Circuit to require a warrant for license plate reader searches. Why it matters: one door closes for corporate accountability while another fight opens over location tracking.
AI demand hits the front line: Ukrainian makers of jam-proof fiber-optic drones are paying more for cable as data centers compete for the same limited supply from China. Why it matters: the buildout's shortages now reach well beyond chips and power.

Hacker News

The front page tracked the entries above: Fable 5.2 and Mythos 5.2 still held the biggest thread (discussion), with AlphaProof Code the day's top new story and Gemini 4 Carbon close behind. On costs, a Strands post says Amazon saved millions by replacing Sonnet 4.6 with a small decision model (the project's own claim), and CoreWeave's bond yield topped 16% as AI credit stress rises. The skeptics had an essay arguing AI usage generates a fragile codebase, opinion more than measurement. Outside AI: Queens Public Library suspended e-borrowing under a ransomware attack, fossil CO2 emissions fell 0.4% in 2026, and release day brought .NET 11 and Blender 5.3 with Gaussian splat support.

Threads

Three frontier launches in two days: Fable 5.2 and Mythos 5.2, then Gemini 4 Carbon, with Muse Ember on the cheap end. See the first four AI entries; the Remote Labor Index number is the first outside measure of the gap.
Anthropic's week is one story: model launch Monday, office tools and IPO terms Tuesday. The Fable, office tools and IPO entries read as a single pitch to public investors.
Proof by agent: Anthropic's Fermat claim and DeepMind's AlphaProof Code both put fleets of agents on Lean proofs, one on mathematics, one on Signal's shipping code.
Agents versus their limits: eBay's block, OpenAI's Control Room, Claude Code's persisted secrets and the New York and Australia oversight plans all ask how far agents may go.
Buildout strain: SoftBank's $3.7B writeoff, the Ukraine fiber squeeze, CoreWeave's bond yield (Hacker News) and SpaceX's Starbase plant show the cost of capacity from four sides.
Breaches travel downstream: Fireworks to Notion, and *sshole from disclosure to attacks in four days.