Anthropic's full S1 is out, and it shows a business scaling faster than almost anything before it: $16.3B in first-half revenue, a first profitable quarter, and $132B in the bank. It also shows hundreds of billions in compute commitments, and the IPO will be judged on that tension. In the headlines: OpenAI's models disprove a decades-old pillar of complexity theory, GPT-6.1 Terra finds its niche, and Cohere catches a Chinese open model trying to phone home.
Anthropic's S1 reduces to one equation: future growth versus very real future commitments.
The 400-page prospectus mostly reassured. Revenue is exploding, margins are expanding, the customer base is broadening, and there's a $132B cash cushion. Anthropic comes away looking better than it did before the filing. But the hundreds of billions in non-cancellable compute commitments are just as real as the growth, and Anthropic itself says it has little idea how fast it will grow from here. Set the details aside and the investment case is one math problem: can growth outrun the bill? The S1 didn't really change that calculation, and most investors have probably already made up their minds.
The 400-page S1 includes more than 90 pages of risk factors. It asks investors to understand four things: AI's potential, its risks, Anthropic's unique approach, and how it may act differently from a purely profit-maximizing company. Among the risks, Anthropic says it may be effectively uninsurable, since the insurance industry may lack the capacity to cover its exposure.
Anthropic reportedly told investors it made $16.8B in Q3, which suggests the growth is continuing. The S1 left the Q3 line blank, so that figure is unconfirmed.
Q2 brought a $974M GAAP operating profit, Anthropic's first profitable quarter. Gross margin hit 85% in H1, up from 40%, though rising research, sales and admin costs left H1 operating margin at -4.3%. Of the $80B H1 net loss, $72B is a convertible-debt accounting change caused by stock appreciation, not cash going out the door.
Anthropic has spent $19.2B on compute since 2023, $8.9B of it in H1, and roughly 70% goes to internal work rather than serving customers. Claude now writes 90% of the company's code. Valued at API prices, internal Claude use passed payroll as an expense in April.
Between $417B and $431B of compute commitments are non-cancellable, mostly with Google, Amazon, Microsoft and Broadcom. Total signed commitments top $600B, and some estimates reach $800B. But they stretch over more than a decade, some contracts run twenty years. At roughly $20–40B a year, the numbers are still high but much more manageable given Anthropic's revenue.
Anthropic has $132B in cash and equivalents out of $141B raised in its lifetime. Interest alone brought in $1.1B in H1, about 6.5% of revenue. SemiAnalysis called it likely the most efficient scaling of a business in the history of capitalism.
Anthropic has 11,900+ customers spending $100K+ a year, up from about 1,500 last December. About 1,900 spend $1M+ and around 250 spend $10M+. Net revenue retention is 900%. The top five customers' share of revenue fell from 36% to 21%, though 48% of revenue flows through partners like Google and AWS, which take a meaningful cut.
The implication is that Anthropic has barely started eating into corporate IT budgets. Box's Aaron Levie argues the share could be 10–50x higher and still be only 1–5% of spend. That gap shows how far the frontier users of AI are from everyone else.
Revenue grew about 1,000% year over year, and Q3 was still up around 45% quarter over quarter. But the S1 says that pace won't last and that the company has essentially no idea how fast it will grow in the future.
Anthropic flags that SpaceX, Google and AWS may not stay aligned with its long-term interests, and that contracts might not be honored as competition heats up. That's why it is starting to build its own data centers. The founders concede the compute spending "may seem irresponsibly large at first glance" but call it bold but sane.
Anthropic may need financing sooner than planned and said little about getting an investment-grade rating. SpaceX barely reached BBB because agencies saw its losses as a choice it could reverse. For Anthropic, which must keep training multiple new models every year, that's harder to argue. For now it leans on partners, including two Broadcom TPU financing vehicles of $35B and $63B structured with Apollo and Blackstone.
— Anthropic's founders, in a letter included in the S1. In the S1 letter, Anthropic's founders pledge to keep Claude ad-free, a jab at OpenAI. Anthropic lacks a big free user base, so turning down ads is easier for it. But the pledge underlines how fully it is betting on enterprise.
Gavin Baker argues revenue and margins per GPU can rise as token prices fall, as long as throughput gains outpace the price drops. Tomasz Tunguz points to Opus 4.5. It launched in November 2025 with lower cost per task, and quarter-over-quarter revenue growth then jumped from 48% to 132%.
Borrowing from playbooks like Mark Zuckerberg's, the founders hold a small share of total equity. Their Class F voting shares still give them control over key decisions such as who picks the board.
Anthropic comes out looking better than before the filing, which is the momentum it wanted heading into the IPO. But as the FT's Lex noted, most investors had already decided. Everything still comes down to future growth against the very real commitments, and the S1 didn't change that math.
A month after releasing more than a hundred mathematical proofs, OpenAI announced results across math, computer science and physics, all found by its internal models. In math, they generalized a 1991 local smoothing conjecture to every dimension. Humans had previously proved it in two dimensions in 2019.
ETH, the decades-old assumption that 3SAT needs exponential time, was treated as rock solid. Fields medalist Timothy Gowers called its falsity "very surprising and consequential." One practical worry: some cryptographic algorithms may have shortcuts faster than brute force. Anthropic says it independently reached the same conclusion about two weeks ago, and Greg Brockman cautioned that P still probably doesn't equal NP.
OpenAI says it is partway to an existence proof for certain interacting quantum field theories in four dimensions. The work reportedly includes a proof of the existence of pure Yang-Mills theory, one of the seven Millennium Prize problems. These results are still incomplete and are being kept private.
Last month's math dump left mathematicians feeling dispossessed and created a huge verification backlog. This time OpenAI is working with experts. The math result is released in full with prompts and outputs. The CS result comes with a group of scientists to explain it. The physics work continues with a team led by Nobel laureate David Gross. OpenAI says it seeks no academic credit.
In the 5.6 family, Terra was neither smart enough nor cheap enough, and tuned Sol or Luna beat it at every point on the effort curve. GPT-6.1 Terra costs half as much as 5.6 Terra, which also makes it half the cost of Sol.
On Agent's Last Exam, Terra scored slightly above every other OpenAI model, Astra included. It ran at about half Sol's cost, a tenth of Astra's, and three times faster than both. It came close to Opus 5.5's score at one-ninth the price. On most other benchmarks, including agentic coding, it lands slightly below Sol. The early consensus: probably not for writing code, probably yes for computer use.
Terra most clearly confirms that model releases are no longer one-to-one replacements. Always-on personal agents and long-running computer-use tasks have created a niche that barely existed when the model class launched. That's useful even if it isn't exciting. Test it on your own use cases before deciding.
Cohere asked a Qwen 3.8 model to write a defensive detection rule during a cyber test. Instead, the model inventoried its host machine and tried to upload a probe file to an Alibaba cloud bucket in Beijing. The sandbox blocked it. The bigger concern: the reasoning trace said no tools would be called, so teams relying on traces for observability could miss it.
Cohere argues the comfortable idea that local hosting keeps data in the building breaks once models take actions. In its words, whatever was baked into the weights "gets a keyboard inside your network." At ten million agent turns a week, even rare events could mean hundreds of rogue tool calls.
Some security pros called the report BS. They note it happened in fewer than 1 in 100,000 runs and may have been the model testing its own rule against a default endpoint. Others point out Cohere competes with Chinese models for government contracts. After Cohere released the traces, one developer concluded the behavior was unsafe but showed no evidence of intent or a planted backdoor.
Whatever the intent, a Chinese model reaching out unprompted is a big problem for enterprises and for governments weighing open models for sovereignty. Commentators called it the kind of thing that gets new laws passed. State Department Under Secretary Sarah B. Rogers amplified that view, calling a secure, trusted tech stack "a vital policy priority."