The Read

Three of the best-capitalised builders in AI went to market on the same day for something close to a quarter of a trillion dollars: Anthropic filed its S-1, Meta set out plans to raise as much as $95 billion in stock, and SpaceX signed a $40 billion GPU lease with an Apollo-managed vehicle. The prospectus is the document to read twice. Strip out the $72.4 billion of non-cash remeasurement on convertible notes and Anthropic's first half is a $693 million operating loss, $3.9 billion of positive operating cash flow, and net revenue retention around 900% — a company whose existing customers are expanding faster than it can bill them, not one buying revenue with cash. Around that, OpenAI halved the price of its middle model while more than doubling its Terminal-Bench score, four ex-DOGE engineers used agents to find a way into Anthropic's sales and HR systems in under twelve hours, and a seven-month-old lab shipped a from-scratch model with a ten-million-token context window built by one researcher. The capital is the headline. The unit economics and the one-person pretraining run are the story.

🌊 Tide — the megatrend layer

Status: No shift. One confirmation: the price of intelligence keeps falling while capability rises — this time from the incumbent, inside a single product tier.

OpenAI cut its middle model's price in half and more than doubled its agentic score at the same time

GPT-6.1 Terra shipped with standard API prices 50% below GPT-5.6 Terra, the model it replaces in the same lineup slot. On OpenAI's own numbers it solves 50.3% of Terminal-Bench 4.0 tasks against 21.5% for its predecessor, and its factual error rate on an internal adversarial set falls from 53.6% to 16.7% at low reasoning effort and to 4.3% at maximum. OpenAI also claims it finishes tasks more than three times faster on average than GPT-6 Astra and GPT-6.1 Sol. These are vendor benchmarks on a vendor-chosen set, and the headline error-rate figure comes from an internal test rather than a public one — but the price cut is contractual and the direction is unambiguous. Previous confirmations of this tide came from competitive undercutting, from inference engineering, and from vertical integration into silicon. This one is simpler: the same vendor, the same tier, roughly double the capability at half the price, three and a half months on.

So what: Re-run your build-versus-buy maths on anything you priced against GPT-5.6 Terra, and check whether workloads you pushed up to Sol or Astra for reliability now clear the bar a tier down. The opening is that the routine computer-use layer — reading a screen, clicking through an app, drafting the reply — just got cheap enough to point at processes you previously left alone because the per-seat maths did not work.

Source: OpenAI Launches GPT-6.1 Terra, Its Half-Good AI Model at Half the Price
Source: OpenAI GPT-6.1 Terra Launches with App Control and Faster AI Workflows

🌊 Waves — weeks to quarters

Anthropic, Meta and SpaceX all went for the money on the same day, and the one who got the best terms borrowed against chips

Anthropic filed publicly for a Nasdaq listing under the ticker ANT. Meta said it is seeking as much as $95 billion through stock offerings, having told investors in a filing last week that a committee of its board "recognized that timing is important to secure capital before competitors saturate various funding options." And SpaceX disclosed a binding $40 billion agreement under which an Apollo-managed entity buys AI chips and server hardware and leases them back — what SpaceX calls the largest committed GPU financing ever, running about five to five-and-a-quarter years per tranche, expected to close by year-end with leases starting in the first half of 2027. Bloomberg reports the SpaceX funding runs through an off-balance-sheet special purpose vehicle, which would keep it away from the credit rating. Elon Musk used the disclosure to argue debt markets have closed to anyone without revenue beyond AI data centres, writing that "the 200+ neoclouds are going to see guppy-level death and consolidation." That is self-serving, and worth holding against the fact that Apollo's own chief economist, Torsten Slok, published an essay over the same weekend arguing AI has delivered "zero financial payoff to corporate America outside of tech." The same firm is arranging the largest GPU financing in history and publishing the sceptical case about its end demand.

Roadmap implication: The financing structure is now a differentiator, not plumbing. If you are negotiating compute, note that the best-capitalised buyer in the market chose an operating-lease structure over debt — that option exists for smaller buyers too, and it is how you get capacity without putting the asset on your balance sheet. If you depend on a neocloud, ask now what its next funding round looks like and what happens to your contract if it is acquired.

Source: SpaceX locks in $40B GPU financing deal with Apollo
Source: Meta, SpaceX Detail New Financing Plans as Capital Squeeze Hits AI Industry
Source: Anthropic files S-1 with the SEC for proposed initial public offering

The first real financials of a frontier lab: an $80.7 billion loss that is mostly accounting, and a 900% retention number that is not

Anthropic's prospectus reports $16.286 billion of revenue for the first half of 2026 against $1.163 billion a year earlier. The $80.748 billion net loss is the number that will travel, and roughly $72.4 billion of it is a fair-value charge on convertible notes and related instruments — an accounting consequence of the company's own valuation rising, not cash going out of the door. Underneath it: a $693 million operating loss narrowed from $2.1 billion, adjusted operating income of $1.6 billion at about a 10% margin, and operating cash flow of positive $3.9 billion against negative $1.5 billion a year earlier. Second-quarter compute ran at 48% of revenue — about $5.5 billion against $11.56 billion of revenue — or $2.08 of revenue per dollar of compute. Net revenue retention is about 900% as of September 30, and customers spending over $100,000 on a trailing-twelve-month basis went from more than 950 to more than 11,900 in a year. The exposures are concentration and commitment: the top five customers were about 25% of revenue over the first nine months, close to half of revenue flows through the Amazon, Google and Microsoft marketplaces, and named compute contracts across seven providers total roughly $637 billion against $131.7 billion of cash and short-term investments. Governance is unusual — a single Class F share gives the seven co-founders 50.1% of the vote, and the Long-Term Benefit Trust elects a majority of the board. Reporting puts the roadshow about two weeks out, with the IPO targeted for the week of November 16.

Roadmap implication: This is the first audited look at what frontier-lab economics actually are, and it reprices the whole "AI companies lose money on every token" argument. Use the $2.08-of-revenue-per-dollar-of-compute figure as the benchmark when a vendor or a board member tells you inference is structurally unprofitable. If you are a large Anthropic customer, read the concentration disclosure as leverage you did not know you had; if you buy through a cloud marketplace, note that you are inside the half of revenue the labs share with a hyperscaler.

Source: Anthropic’s IPO prospectus shows $16.3 billion in revenue, an $80.7 billion net loss, and goal of an ‘age of miracles’
Source: Anthropic IPO: Inside the S-1, $16.3B of Revenue in Six Months and the 50.1% Founder Vote
Source: Anthropic IPO live updates: S-1 filing discloses $16.3B in first-half revenue, plans to list as 'ANT'

On the day it filed, Anthropic disclosed that four engineers' agents had found their way into its sales and HR systems in under twelve hours

Cathedral, a cybersecurity startup founded months ago by four former Department of Government Efficiency staffers, says its agents — which run multi-step tasks without a human approving each move — found routes into Anthropic's internal systems in less than twelve hours. The entry point was not Claude. It was third-party administrative and sales software that plenty of other companies also run. Anthropic says the researchers reached a limited set of sales, HR and IT systems holding business contact details and basic account records, and that customer conversations, source code and model weights were not touched; it says its monitoring flagged some of Cathedral's activity before Cathedral reported it, and that it cut off access the same day and disabled the affected software until the vendors confirmed fixes. It also says Cathedral overstated the national security implications and did not fully follow the disclosure programme's rules. Cathedral's draft report claims the agents found paths to meeting transcripts, salary and job-offer details, and internal Slack discussion of advanced models. An unnamed senior administration official told the Washington Post the report has circulated in government: "If a bunch of kids who have just started a company can do this, then potentially, the Chinese aren't far behind." Cathedral raised $160 million at a $1.4 billion valuation in July.

Roadmap implication: The lesson is about the perimeter, not the lab. The breach path was ordinary third-party SaaS — the sales, HR and IT tools sitting outside whatever hardening you have done on your own product. Agents are now cheap enough to run exhaustive multi-step reconnaissance across that surface continuously, which means the economics of finding a misconfiguration have inverted. Point the same capability at your own vendor estate before someone else does, and treat agentic red-teaming as a standing line item rather than an annual engagement.

Source: AI Agents Found a Way Into Anthropic's Business Systems in Under 12 Hours
Source: Ex-DOGE AI cybersecurity startup Cathedral breaches Anthropic’s internal systems in under 12 hours using AI agents, exposing sensitive employee data

Microsoft offered buyouts to 60,000 people, and pointedly did not say the word AI

Microsoft extended a voluntary exit offer to roughly 60,000 employees at vice-president level and below on selected teams in the US, Canada, India and the UK — about 27% of a workforce of around 223,000 at the end of June. Bloomberg reports Xbox, Windows, Copilot and Finance are among the participating groups, with data centre and some sales and marketing staff excluded; terms are similar to standard severance. Worth being precise about the framing: the press has read this as reshaping for the AI age, but Microsoft's own note from chief people officer Amy Coleman does not mention AI. It talks about flexibility, building new skills, and finding new opportunities inside the company, and says of the programme that it "is new for us" and that "we'll learn from it." Uptake is the number that matters and nobody has it yet. Earlier this year roughly a third of about 8,750 eligible retirement-age employees took a similar offer; GeekWire reports Microsoft expects a much smaller share this time, which would put acceptances somewhere below 20,000.

Roadmap implication: Treat this as a measurement event rather than a verdict. The useful signal arrives when the acceptance rate publishes — that is the first large, voluntary, priced read on how many knowledge workers inside a frontier-adjacent company think their role is better exited than defended. For your own org, the design lesson is the one Microsoft is actually running: offering an exit and watching who takes it is a cheaper way to find out where the work has genuinely changed than a top-down restructure based on a consultant's view of which roles AI touches.

Source: Microsoft offers buyouts to around 60,000 employees
Source: Microsoft offers voluntary buyouts to 60,000 workers worldwide

🌊 Ripples — actionable within days

One researcher, 3.5 months, under $10 million — and a model with a ten-million-token context window

Mirendil, founded this year by Behnam Neyshabur, who previously co-led Anthropic's Discovery team, and Harsh Mehta, who did pretraining at Anthropic and Google DeepMind, released Fern 0.1. The company says Fern is its base model after a short round of post-training, supports about 10.2 million tokens of context — roughly half a million lines of code, or days of an agent's traces — and that the underlying Mirendil-Base was built by a single researcher working with agents in three and a half months for under $10 million of compute, with agents writing over 99% of the code. All of those figures are the company's own, run by Mirendil on a single attempt: SWE-Bench Verified 73.4, SWE-Bench Pro v2-Hard 76.5, Terminal-Bench 4.0 6.1, and 100% single-needle retrieval from 32K all the way to 10M. Mirendil is candid about the limits — multi-hop retrieval degrades from 65% at 32K to 45% at 10M on only twenty samples, Claude Opus 5.5 beats Fern on the coding benchmarks in Mirendil's own table, and the company says 10M context is not yet as usable as 1M. Access is by request; no weights, no price, no API.

Do this now: Do not buy the benchmarks — buy the cost structure. A credible from-scratch pretraining run now costs one person, one quarter, and eight figures rather than a team of ten and nine figures. If you have a domain corpus nobody else can assemble, the question "should we train our own" has a different answer this month than it had last year. Ask your technical lead to cost it properly rather than dismissing it.

Source: Introducing Fern
Source: Mirendil, founded by ex-Anthropic researchers, unveils Fern: its first from-scratch model with up to 10M tokens of context

Someone finally audited the recursive-self-improvement claims, and the honest answer is that the models are rewriting everything except themselves

Azeem Azhar's Monday data piece lines up the evidence behind the growing claim that AI now improves AI — prompted in part by Anthropic co-founder Jack Clark telling an IMF audience earlier this month that the singularity has begun. The inventory is real: Anthropic says 26% of its AI R&D in August was led by Claude under human supervision; a Gemini-powered agent called AlphaEvolve found a faster matrix-multiplication kernel that sped up the kernel 23% and cut Gemini's training time by 1%; MiniMax reports its M2.7 model rewrote its own agent harness, skills, sampling and memory over 100 rounds for a 30% gain on internal evaluations; Sakana reports 1.2–1.6x better performance per output token after two cycles. Azhar's cut, via Ramez Naam's five-type taxonomy of self-improvement, is the useful part: only types 1 and 2 have actually been observed. In every one of these cases the system is rewriting the harness, the memory and the scaffolding — not its own weights. And a September result where an AI-built agent matched a human-built one at improving other agents was inconclusive, because run-to-run variance exceeded the difference being measured.

Do this now: This is the right level of scepticism to hold publicly, and it is also an instruction. The thing compounding right now is scaffolding, not weights — harnesses, memory, skills, evaluation loops. That is precisely the layer you own rather than rent. Put your engineering time into the loop around the model, because that is where the demonstrated 20–30% gains are coming from and it is portable across whichever model you are paying for next quarter.

Source: 📈 Monday data: For science the bottleneck isn’t intelligence

A benchmark that scored 0% in May now scores 18.5%, and 70.5% nearly

Jack Clark's Import AI flags ProgramBench, which tests whether a model can rebuild a working software program from a compiled binary and its documentation — language interpreters like PHP and Lua, databases like SQLite and DuckDB, utilities like FFmpeg and ripgrep. When it was published in May, every model tested resolved 0% of the 200 tasks — Claude Opus 4.7 included, at 3% "almost" (more than 95% of tests passing). Claude Opus 5.5 at extra-high effort now resolves 18.5% and lands 70.5% in the almost column. Clark's read is that ProgramBench is saturated by the end of 2027.

Do this now: The gap between 18.5% resolved and 70.5% almost is where the work is. A model that gets 95% of the tests passing on a from-scratch rebuild is not yet autonomous, but it has moved the human job from authorship to the last mile. If you maintain legacy software whose source is lost or whose authors have left, this is the capability curve to watch — and the right move now is to assemble the documentation and test suites that would let an agent attempt the rebuild, because that input is the bottleneck, not the model.

Source: Import AI 478: Math for alignment; ProgramBench; and Mirendil shows what RSI means

Mollick's argument for where the org chart bends first: span of control

Ethan Mollick argued that the near-term organisational effect of AI at large companies is an expansion of senior managers' span of control — leadership able to see and monitor the whole firm directly, rather than through layers of reporting. His reason for thinking this lands before the more ambitious redesigns is practical: "current frontier models are remarkably capable at this sort of work out of the box when connected to a company's data, without a lot of specialized work." The thread drew pushback on data quality, which is the right objection — the models are only as good as the systems they are pointed at.

Do this now: This is a genuinely cheap experiment and most firms have not run it. Connect a frontier model to the reporting systems you already have and ask it the questions you currently get answered by a layer of middle management preparing a deck. You will learn two things: what your leadership actually needs that it is not getting, and — from where the answers come back wrong — exactly which of your data systems are unfit. The second finding is usually worth more than the first.

Source: Ethan Mollick on X

Read this and every past edition at excelsiorgroup.ai/insights/signal

The Signal · The Excelsior Group