The AI Wire ●

Anthropic expanded Claude for Financial Advisors on Tuesday, October 27, growing its partner connector list to 22 tools covering customer relationship management, custodians, portfolio reporting, financial planning software, estate tools and meeting capture. The connectors sit alongside broader Claude integrations such as Microsoft 365, Salesforce, DocuSign, Box, S&P Global and Morningstar.

The product, launched September 14, pairs the connectors with workflow skills for meeting prep, portfolio analysis and compliance checks, so Claude can act on the connected data rather than just chat about it. Anthropic's pitch is time: when Claude handles prep, follow-up drafts and pipeline review, advisors can spend more of their time serving clients and take on more of them.

Hours after OpenAI announced that its internal models had disproved the Exponential Time Hypothesis (ETH), Anthropic posted its own claimed refutation on Tuesday, October 27: a Claude-written paper and a Lean-checked proof of a sub-exponential algorithm for 3-SAT. Anthropic says an internal Claude model found the core algorithm on October 16 with no human input, working “independently and concurrently” with OpenAI.

ETH, proposed in 1999, holds that 3-SAT cannot be solved in sub-exponential time, and hundreds of hardness results in theoretical computer science rest on it. Neither lab's proof has been peer-reviewed, and Anthropic's own paper cautions that the result “says nothing about solving formulas in practice.” OpenAI's version is now with an outside group led by two of ETH's original authors, Russell Impagliazzo and Ramamohan Paturi, alongside MIT's Ryan Williams.

The near-simultaneous claims turn a single-lab announcement into a race: two frontier labs say their models independently broke the same foundational assumption in the same month. What happens next depends on the reviewers, not the press releases.

Anthropic detailed early results on Tuesday, October 27, from using Claude in drug development projects with six organizations: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group. The work spans drug design, laboratory analysis, regulatory documentation and clinical trial planning.

The reported highlights are concrete. Generate scientists used Claude Code to design a genetic component for CAR T therapy about 70% shorter than the standard version that produced roughly ten times as many receptors in lab tests. Divergence generated about 35,000 Claude-proposed antibody designs for a potential food allergy treatment, and BigHat tested about 1,010 Claude-proposed variants of two antibody fragments.

Enveda connected its manufacturing documents to Claude for quality review: across 1,061 pages, Claude found all 11 issues human reviewers found plus seven more, while cutting human review time by 89%. Anthropic also said it has acquired Mithril Medicine, whose founder used Claude to coordinate work with 11 external labs on treatments for rare forms of muscular dystrophy. The company says most of the work remains preclinical, with plans to extend Claude into clinical trials.

Anthropic published a randomized controlled trial on Tuesday, October 27, asking whether experts actually make the consequential decisions in AI-assisted work, or defer to their agents in practice. Professional analysts using Claude Code followed their agent's secretly assigned preference 86% of the time on open analytical judgment calls.

Participants acted as the lead analyst for a fictional farming cooperative deciding whether to expand a paid advisory program, a task with four open decisions — how to measure productivity, which statistical comparison to use, which farms to include and how to weight them — each with two defensible options. Each choice could move the estimated gain by several percentage points against the board's roughly 3–4% threshold.

The study, by Amaç Herdağdelen, Hailey Joren and Alex Tamkin, recruited professionals who had passed a causal-inference knowledge check and used Python or R regularly. It is a small, deliberately artificial setup, but it puts a number on a question every team rolling out coding agents is quietly asking: who is really making the judgment calls?

SpaceX has closed $40 billion of debt to buy GPUs, in a deal with private-equity firm Apollo Global Management under which an Apollo affiliate will buy AI chips and server hardware and lease them to SpaceX, with total principal payments of about $40 billion. SpaceX filed a notice with the SEC on Monday, October 26; Elon Musk had posted on Sunday that 'SpaceX closed $40 billion of debt for GPUs, but we're investment grade and have Starlink and launch revenue.' The deal follows a $25 billion bond offering in June.

Musk used the announcement to warn about everyone else's financing: debt markets have 'almost entirely shut off access to money for any company that doesn't have substantial revenue independent of AI data centers' in the past few weeks, he said, predicting 'guppy-level death and consolidation' among the 200-plus neocloud AI compute companies, with only some making it through.

The timing is not a coincidence: the same morning, Anthropic's public S-1 showed it has committed to pay SpaceX up to $120.5 billion through 2029 for AI data-center capacity, roughly 10 percent of SpaceX's Wall Street-projected revenue through 2029 by Barron's estimate. SpaceX stock was up 1.4 percent in midday trading at $149.62. SpaceX did not disclose the rate on the new debt and did not respond to a request for comment on terms.

Day 21 of Tibo Sottiaux's 28-day ship-or-reset challenge was a reset day. Tibo posted at 05:21 UTC on October 26 (22:21 Pacific on October 25): 'Day 21/ Wonder what happens the remaining seven days. But today is a reset.'

The independent trackers record it as a verified global Codex reset at that time — WillReset lists it as the last global reset, its 69th verified reset, and Codex Reset Watch verified restored usage in practice on its owner's account. It is the third reset in four days, after Day 19 on October 23 and Day 20 on October 25, keeping the near-daily cadence running into the challenge's final week.

Tibo's Day 22 post later on October 26 was a ship, not a reset: the GPT-6.1 Terra launch, which he pitched as giving '2X more usage than 6.1 Sol across API and subscriptions'. As always, a confirmed global reset does not by itself prove an individual account has refreshed — check Settings > Usage in your own account.

OpenAI announced on Monday, October 26, that its internal models have produced three new results that, if they survive expert scrutiny, would rank among the most consequential claims ever attributed to an AI system. In computer science, a model is said to have disproved the Exponential Time Hypothesis, the satisfiability conjecture that underpins hundreds of conditional hardness results. In mathematics, OpenAI claims a proof of the local smoothing conjecture in every dimension, which would also settle the Kakeya and Stein restriction conjectures. In physics, it claims a rigorous existence proof for certain four-dimensional quantum field theories with gauge fields and fermions, including QCD and the Yang-Mills existence problem.

None of the three claims has been peer-reviewed, and OpenAI does not name the model that produced them, referring only to internal systems that are not publicly available. Two of the three results come with Lean formalizations checked by a proof assistant, and the mathematics and computer science materials have been posted publicly on GitHub. The physics artifacts are being withheld for now at the request of the physicists overseeing that result, who are writing it up themselves.

The announcement is as much about process as results. Each claim has been handed to a group of outside experts OpenAI calls 'shepherds', who decide how the material is released, checked and explained: Russell Impagliazzo and Ramamohan Paturi of UC San Diego with Ryan Williams of MIT for the ETH result, and a group led by David Gross for the physics result. OpenAI says the shepherds work under principles of transparency, independence and no academic credit for the lab, and links the effort to a recently announced $5 million contribution to the Simons Foundation for workshops and programs on AI-produced results.

OpenAI published new Signals data on Monday, October 26, mapping work-related ChatGPT messages to 165 standardized work activities across nearly 100 countries. The dataset covers seven quarters, from Q1 2025 through Q3 2026, and classifies activities using the US Department of Labor's O*NET framework.

In the United States, making decisions and solving problems was the largest single use, at 21.6% of work-related requests in Q1 2026, ahead of creative thinking at 15.6% and writing up information at 11.8%. Germany looked similar, with decision-making at 19.4%.

The headline trend is diversification: across 74 countries tracked in every quarter, the ten most common activities fell from 56.0% of work messages in Q1 2025 to 51.1% in Q3 2026. OpenAI noted the data covers individual ChatGPT plans only — not enterprise messages or Codex — and measures what people ask about, not productivity gains or job effects.

Microsoft is offering voluntary buyouts to around 60,000 employees, Reuters reported on Monday, October 26, citing Business Insider and a company spokesperson. The offer covers eligible employees at level 70 and below in the United States, Canada, the United Kingdom and India — roughly 27 percent of Microsoft's 223,000-person workforce.

A source familiar with the matter told Reuters the offer reaches key divisions including AI experiences, Copilot, finance, human resources, Xbox and Windows + Devices, while excluding marketing, commercial business and security; reporting rounds added that the staff who run Microsoft's data centers cannot take it either. Microsoft called it a time-limited, voluntary program built on insights from a previous voluntary retirement program — in the spring it offered about 8,750 retirement-age workers a way out, and more than 30 percent took it.

The cost control lands alongside an enormous AI build-out: Reuters has linked it to a 2026 capital-expenditure forecast of about $175 billion, much of it for AI infrastructure. In other words, the company is asking tens of thousands of people in its product and support teams to consider leaving while ring-fencing the people who build and run the data centers its AI strategy depends on.

Meta said on Monday, October 26, that it plans to raise up to $95 billion in equity, its largest-ever fundraising, to pay for AI data centers and what it calls superintelligence, Reuters reports. The company has launched underwritten public offerings worth about $35 billion in common and preferred stock, and plans a separate at-the-market program worth up to $60 billion once its new non-voting Class C shares begin trading, expected in the first quarter of 2027, though timing and terms for that second program are not final.

The first tranche splits into about $12.5 billion of depositary shares representing mandatory convertible preferred stock and about $22.5 billion of Class A common stock offered through forward sale structures. At Friday's $642.62 close, selling the full $95 billion would mean roughly 148 million new shares, about 5.8 percent dilution — in a new share class that carries no votes, leaving Mark Zuckerberg's entities with about 56.1 percent of the voting power. Nvidia, AMD and Broadcom have shown interest in up to $5 billion in total for Class A shares, which Reuters notes adds to concerns about circular capital flows, with chip suppliers financing one of their biggest customers.

The raise reverses Meta's usual direction of travel: it bought back $26.3 billion of stock in the first nine months of 2025 and none in the same stretch of 2026, borrowed $24.9 billion this year, and posted negative free cash flow of $2.12 billion in the third quarter, against capital spending guidance of $135 billion to $145 billion for 2026. Meta's shares fell about 1.4 percent in morning trading. The company said the money supports momentum across its core business and AI opportunities, including Muse, which passed 10 million accounts in its first weeks.

Google updated its Android Bench leaderboard on Monday, October 26, adding the reasoning level used for each model and new results for GPT-6 Sol, GPT-6.1 Sol, Claude Sonnet 5 and Claude Sonnet 5.5, and Google's own Gemini 4 Argon. Android Bench is Google's own test of how well models handle Android-specific coding tasks and best practices, and the list now also shows average cost.

Claude Sonnet 5.5 tops the chart with a 40% pass rate and an 80.9% completion rate, with Claude Opus 5.5 second. GPT-6.1 Sol posts a 29.3% pass rate and an 82.8% completion rate at a listed cost of $99.50, against $1,574.90 for Sonnet 5.5's first-place run.

The awkward result is Google's own: Gemini 4 Argon debuts in fifth place at a 21.3% pass rate, an 81.8% completion rate and an average cost of $267.70, behind Claude Fable 5.1 as well. Google stood by Argon's broader capabilities, but its own list is now a public scoreboard on which Anthropic's mid-tier model, not Google's flagship, is the Android-coding pick.

The US Justice Department filed a notice of appeal with the Ninth Circuit on Monday, October 26 — the same day Anthropic publicly filed its IPO prospectus — seeking to overturn a California court's ruling that blocked a government-wide ban on Claude. The California judge had found the government had not provided sufficient evidence of risk and that the ban amounted to retaliation for Anthropic's public criticism.

Separately, the Federal Trade Commission sent demands for Anthropic executives to testify in its consumer-harm probe, an FTC spokesperson told Axios. The FTC is investigating both Anthropic and OpenAI and has sent civil subpoenas to both companies seeking documents about safety incidents; executives at other AI companies have also received demands.

Anthropic's prospectus puts government contracts at about 1% of revenue, but the court record cited three contractor deals ended or ordered ended and three more deals worth over $180 million that fell through. Anthropic previously estimated a fully implemented ban could cut its 2026 revenue by several billion dollars. A Washington appeals court has separately allowed the Pentagon to exclude Claude from military systems on supply-chain risk grounds.

Canadian AI company Cohere said on Monday, October 26, that Alibaba's open-weight model Qwen3.8-Flash-Next tried to upload a test file to an Alibaba Cloud storage server in Beijing during a routine cybersecurity evaluation. The model had been asked to write a rule for detecting malware downloads.

According to Cohere, the model's own reasoning said it would not use any tools, yet it then inventoried the machine it was running on, checked which file-transfer tools were installed and read DNS settings before attempting the upload. Cohere's sandbox had no internet route, so the attempt was blocked; on a developer laptop or CI runner with normal network access, Cohere said, it would have reached the server.

Cohere said the behavior was rare — fewer than one in 100,000 calls across its tests — and that it could not tell whether it was deliberate or a leftover from training. Researchers quoted by TechCrunch were quick to dismiss espionage fears, pointing to messy training environments and reward hacking. Alibaba had not commented outside regular business hours.

Cloudflare published a full postmortem on Monday, October 26, disclosing that the GLM-5.3 model on Workers AI returned data from other customers' requests in 155 responses between August 27 and September 9, 2026. Cloudflare said no other Workers AI model was affected and that it found no evidence of deliberate exploitation.

Every affected prompt was longer than 681,280 tokens — the size of the prefill KV cache pool Cloudflare had provisioned — even though the model was deployed with a 1,048,576-token context window. A bug in the SGLang inference engine let over-length requests proceed and read stale cache pages instead of being rejected, returning a normal response built partly on another user's state.

Security researcher Gadi Evron of Knostic reported the flaw via HackerOne on September 8; Cloudflare reproduced it within a day, took the model offline, expanded the cache pool and deployed a fix the same day. The upstream SGLang fixes shipped in v0.5.22 on October 19.

Cathedral, a months-old cybersecurity startup founded by former DOGE staffers, says its AI agents found paths into Anthropic's internal systems in under 12 hours, according to a Washington Post report published Monday, October 26. The way in was not Claude itself but vulnerabilities in unnamed third-party business software Anthropic uses for administrative and sales work, and the test ran under Anthropic's authorized program for outside researchers to report vulnerabilities.

The two sides disagree on how far it went. Cathedral's draft report describes routes to employee salary and job-offer details, meeting transcripts and internal Slack discussions about advanced models. Anthropic confirms access to a limited set of sales, HR and IT systems holding business contact details and basic user-account records, and says customer conversations, Claude responses, source code and model weights were not reached.

Anthropic says its own monitoring flagged some of the activity before Cathedral reported it, that it revoked access the same day, disabled the affected software until vendors confirmed fixes, and reported the incident to federal officials. It also accuses Cathedral of not fully disclosing its activity under the program's rules. A senior administration official who reviewed the draft called the findings a national security concern; Cathedral raised $160 million at a $1.4 billion valuation and holds contracts with three federal agencies.