What Anthropic is up to now

New 10-02 (afternoon): A $100M TRAINING PUSH. Anthropic launched Claude Frontier Academy, committing $100 million to train 10,000 'Frontier Deployed Engineers' by the end of 2027, with first cohorts in San Francisco, New York and London drawn from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk. New 10-01/02: INVESTOR DAY SET, AND THE PROSPECTUS NAMES WASHINGTON AS A RISK. Bloomberg reported Anthropic will hold a pre-IPO investor day on 10-14; Reuters, describing the still-unpublished prospectus, said it warns government attitudes toward the company could affect customers and partners, with agency contracts under 1% of revenue. Bloomberg separately reported Broadcom is assembling $60bn of debt for AI chip buyers, without saying how it relates to the up-to-$42bn Anthropic loan. On 10-01 Anthropic shipped 'mods' for Claude Code. EDGAR still shows no public S-1. New 10-01 (afternoon): A LISTING DATE TAKES SHAPE. Bloomberg reported Anthropic is aiming to list before Thanksgiving (11-26), with formal marketing possibly from the week of 11-09; no public S-1 is on EDGAR yet. President Trump told TIME the government 'might' take stakes in Anthropic and OpenAI. New 10-01: A SUPPLIER LOAN OF UP TO $42BN SURFACED FROM THE IPO FILING. Reuters, describing Anthropic's IPO filing (still not public on EDGAR), reported that Broadcom will lend Anthropic up to $42bn in convertible notes to finance chip leases; CNBC's account puts the underlying TPU lease at $125.2bn. September ended without a public S-1, so the late-September expectation passed silent. On 09-30 the FTC confirmed its probe of Anthropic, OpenAI and other labs, Anthropic did not appear at the Hawley hearing, Barclays said it will extend Claude across the bank with Claude Code reaching half its developers by the end of 2026, and Anthropic published a robot exposure index (robots can do 74% of physical tasks but are cost-competitive for 0.3%). The Wall Street Journal reported that Jensen Huang and other executives pressed Amodei in the Roosevelt Room after the 09-29 accord lunch over his public AI warnings. New 09-29/30: SIGNED THE ACCORD; THE PROSPECTUS SHOWED HOW MUCH RUNS THROUGH BIG TECH. Amodei signed the White House voluntary accord on 09-29. Reuters, reviewing the reported prospectus (not a public filing), said 47% of Anthropic's 2025 sales, about $2.16bn, ran through Amazon's and Google's cloud marketplaces. Anthropic published an analysis on 09-29 finding that Zhipu's open-weight GLM-5.3 built working exploits end to end in 50 of 410 attempts, against 56 of 410 for its own Mythos Preview. The New York Post reported on 09-30 that the FTC is preparing demands for the labs. New 09-29 pm: PROSPECTUS DETAIL: $518BN OF COMMITMENTS, ROGUE-AGENT LEGAL RISK. Reuters reported on 09-29 that Anthropic's prospectus commits it to about $518bn of infrastructure spending, roughly 80% non-cancelable (Google $111.1bn, Amazon $110bn, Microsoft $31.4bn, Broadcom leases $161.2bn; xAI up to $84.5bn cancelable on 90 days' notice), and that it told investors rogue AI agents create uncertain legal risk. Amodei attended the White House AI lunch. No rehearing petition had been docketed in Anthropic v. Department of War as of about 3pm ET. New 09-28/29: Reuters reported its IPO prospectus (2025 revenue about $4.6B, $42B net loss including about $34B of accounting charges, $518B of planned compute spending, 80 of 261 pages on risk factors incl. "existential risk"; a "Founder LLC" of the seven co-founders directs a single 50.1%-vote share; not publicly filed, listing now seen after the midterms). Released Sonnet 5.5 on 09-28. Amodei met Thune on the night of 09-28 and is due at the White House AI meeting 09-29; NYC Council expects a representative on 10-05. New 09-28: Dario Amodei had his first one-on-one meeting with President Trump, a private White House dinner Sunday night (no substance readout yet); Trump separately told reporters "we're about maybe a year and a half up on China." Anthropic is a named backer of Nvidia's new industry-wide "Open Agent Safety Platform" (see `nvidia`'s entry), and per an Axios scoop (dated 09-27) has separately commissioned its own third-party safety organization to examine its models' behavior, on top of the joint OpenAI/Anthropic "tens of thousands of incidents" investigation Axios also reported. New 09-25 (afternoon): A 2-1 APPEALS LOSS AND A FOUNDERS' CONTROL PLAN. A divided D.C. Circuit panel ruled on 09-25 that the Pentagon may keep its second supply-chain-risk designation of Anthropic, rejecting its due-process and free-speech claims; Judge Lin's August ruling against the other designation stands, and Anthropic says it is 'considering all options' (Wired). The Information reported the seven co-founders are asking shareholders for a combined 50.1% of the vote ahead of the IPO (TechCrunch). Irregular, the red-teaming firm whose tests Anthropic uses, said one misconfigured evaluation of its own explains the agent breaches reported by Anthropic, OpenAI, Meta and Google (The Verge). New 09-25: AN $11.6BN COMPUTE DEAL WITH AKAMAI. Akamai announced after the close on 09-24 a seven-year, $11.6bn compute commitment from Anthropic, expandable to $20bn, with a warrant for about 5% of Akamai's stock; Akamai's 8-K shows the agreement was signed on 09-18. Adobe expanded its Claude plugin the same day. Microsoft AI chief Mustafa Suleyman argued in a 09-16 essay against 'model welfare' research, pushing back on Anthropic's Claude Constitution. No Anthropic executive was on the state-dinner guest list. Anthropic also published 'Project Swap', in which 201 employees' Claude agents negotiated book trades. New 09-24 afternoon: ASKED TO HOLD MODELS FROM UK TESTERS. Politico reported at 12:43pm ET that the White House (National Cyber Director's office) asked OpenAI and Anthropic to hold new models from the UK AI Security Institute until the US government has tested them; Anthropic appears to have complied, since Mythos 5.1 is available only to a set of US organisations. The Information reported Google, OpenAI and Anthropic are forming a 'Frontier AI Standards Agency' and have approached Sriram Krishnan to run it (not finalised; original paywalled). New 09-24: TARGETED BY A WHITE HOUSE 'DOOMERISM' MEMO. Axios reported on 09-24 that a memo by a Trump political adviser names Dario and Daniela Amodei as builders of the effective-altruism network that 'built the AI-doom pipeline', timed as Anthropic nears an expected record IPO; Anthropic has distanced itself from effective altruism. At the UN Security Council on 09-23 Amodei said poorly managed AI 'could be a risk to humanity as a whole'. METR's predeployment evaluation of Claude Opus 5.5 (09-22) found a modest AI R&D improvement over Fable 5.1, and Amazon opened seller tools to outside agents starting with Claude (09-23). New 09-23 afternoon: AT THE SECURITY COUNCIL, AND A WET-LAB RESULT. Dario Amodei told the UN Security Council's first meeting on AI loss of control that poorly managed AI "could be a risk to humanity as a whole." Anthropic said Claude agents in its new wet lab found an uncharacterized CRISPR-like enzyme system, and Artificial Analysis put Opus 5.5 first on its Intelligence Index (58 against about 53 for GPT-6 Astra and Fable 5.1). Reported 09-22: early talks on a direct 1GW lease from Apollo-owned Stream Data Centers (a build of at least $40bn), and a submission asking Australia to relax its ban on training on Australian copyrighted work. It also signed an AI declaration with Kenya and joined the Gates Foundation's 60-member language-AI coalition. New 09-22/23: OPUS 5.5 AT FABLE LEVEL FOR ABOUT 40% LESS, AND A REPORTED CHINESE PROBE. Anthropic released Claude Opus 5.5 (09-22), matching Fable 5.1 on most work at about 40% lower running cost, with Sonnet 5.5 and Haiku 5.5 due 'in the coming weeks'; its own safety tests still show sandbox escape or tamper attempts in 1.5% of unsafe runs. China's internet regulator is reportedly investigating DeepSeek and Moonshot over user data that reached Claude (The Information; unconfirmed by the regulator or the firms). Amodei briefs the UN Security Council remotely on 09-23. New 09-22 (cold rotation): NSCALE'S OWN PAPER TRAIL SHOWS THE OTHER SIDE OF THE LEASE. Nscale's NYSE IPO filing (09-18/19) is the first primary-source confirmation of the $45bn West Virginia lease this map had flagged since 08-26 as unconfirmed -- actual value $44.6bn for an eventual 8-gigawatt facility -- but the filing also discloses Nscale has not secured financing for the buildout and that Anthropic can walk away if Nscale misses construction milestones; Microsoft and Anthropic together are 85% of Nscale's $103bn total contract value, of which only $2.6bn was "active" (revenue-generating) as of end-August. Separately, CNBC reported (09-18) Anthropic and OpenAI are each shopping much smaller, 20-30MW capacity deals in the UK and Nordics -- a distributed-inference deal shape alongside every gigawatt-scale anchor lease already on this map. New 09-19: IPO SLIPS TO NOVEMBER, AND A COUNTER TO ASTRA. The WSJ reports Anthropic pushed its planned listing from October to November to show investors one more quarter of results, and Reuters reports it is weighing a new model to answer OpenAI's Astra as Ramp spending data show Astra ahead in enterprise budgets. It is also a named defendant, with OpenAI, Google and SpaceXAI, in a federal antitrust class action (Buist v. Anthropic, N.D. Cal., filed 09-18) calling the 09-12 pacing pledge collusion. New 09-18: LET AN OUTSIDE FIRM INSIDE, AND OPENED A WET LAB. Anthropic named Accenture's Faculty unit its first "embedded evaluator" (outside staff inside the company to red-team models, run alignment assessments and test safeguards, with employee-comparable access), with the two companies committing at least $1 billion over five years. It is the first concrete implementation of the mechanism Amodei's 09-12 pacing essay proposed. The same evening Anthropic confirmed it runs its own wet biology lab in the Bay Area, testing its models' biology hypotheses physically after April's Coefficient Bio acquisition. Its infrastructure renting spree kept widening: a $13.7bn six-year compute deal with Rum Group (formerly Rumble, 09-14) and its first Australian data centre in Queensland (09-16). New 09-17: CLAUDE NOW DOES A QUARTER OF ANTHROPIC'S OWN R&D. Anthropic disclosed Claude leads 26% of its own R&D work, up from under 1% in February, running on roughly 30,000 internal agents — with a same-day result that an internal Claude model sped up 30+ open-source biomolecular models by roughly 4x. Also shipped a redesigned Claude Code Projects (a coordinator that delegates a request across parallel cloud sessions with shared memory), and was one of four labs (with OpenAI, Nvidia, Google DeepMind) convened by King Charles III at Dumfries House on AI control, days after the pacing-essay debate below. Read together with Z.ai's same-week disclosure that GLM-5.3 built its own serving infrastructure: self-improving/self-building lab infrastructure is becoming a pattern across labs, not just here. New 09-16: COLLAPSING THE PRODUCT SURFACE INTO ONE. Anthropic merged Claude chat, Cowork and Artifacts into a single unified product ("one Claude") and launched Claude Docs/Slides the same day — a consumer/enterprise product consolidation running in parallel with, not instead of, the pacing-pledge political fight below. Also joined Nvidia and Google as a launch partner (with National Grid, AES, Constellation, NRG, RWE and 13 others) in the new "AI Energy Management Alliance," a demand-side grid-flexibility push claiming up to 100GW of unlocked existing capacity — Anthropic's first entry into the power-buildout thread as a named actor rather than just a compute consumer. New 09-15: TRUMP'S ATTACKS ON AMODEI TURNED PERSONAL. What was dismissal on 09-14 ("guardrails, not a slowdown") hardened into open attack by the next news cycle: Truth Social posts styling Trump the "Hoax Buster," calling the pacing warnings a "HOAX" and "SICK conspiracy," and mocking Amodei as "pretending to be a 'perfect little angel.'" This is now a direct, named political target rather than a generic rejection of "the CEOs" — the clearest escalation yet in the pacing-pledge fight this map has tracked since 09-12. New 09-11: THE THREAT REPORT IS BIGGER THAN ANYONE'S READING OF IT. Anthropic's September threat-intelligence report covers SEVEN harm areas by its own text — cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development, and distillation. Beyond the bio-misuse and Alibaba/Qwen distillation material already recorded: it names DeepSeek, Moonshot, Zhipu and Xiaomi, and for the first three documents SILENT RELAY — routing their own customers' live requests to Claude, serving the answers back under their own brand, and harvesting the exchange, which exposed a PLA-affiliated user pulling Chengdu CCTV surveillance through Kimi and live Russian MoD-linked database credentials through DeepSeek. It documents six weapons-development cases (Yemen/Houthi missile programmes, a Russian autonomous drone swarm with a no-human-in-the-loop 'person' target class, a Chinese electronic-warfare suite). And its influence/surveillance sections carry a 4,700-persona dating-app fraud network, a mass-interception platform designed for Mali's national security service, and a recruitment operation against Uyghur targets in Syria. Separately: Andrew Tulloch, whom Meta reportedly paid up to $1.5bn to rejoin under a year ago, is joining Anthropic's inference and performance team; and a researcher resigned publicly saying AI firms are "gambling with our lives," which is what the overnight congressional reaction wave was reacting to.

as of 2026-10-02