In early November, Moonshot AI released the weights of Kimi K3.1 Pro, a 2,800-billion-parameter model that any company can now download, within the limits of a licence governing commercial use. On Artificial Analysis's index, it scores 46 points, the best score for open-weight models. On 5 November, evaluation body METR published a report saying Anthropic's internal models have been progressing about twice as fast since October 2025. Four days later, Anthropic released Claude Fable 5.2: 65 points. On Wednesday 18 November, the company must set its IPO price, with $900.9 billion in compute and infrastructure commitments in its prospectus. Do these billions buy a lasting lead, or just a few months?
What the measures say
The best open models reach the level of closed ones with an average four-month lag (Epoch AI, May 2026), four to seven months on cybersecurity tasks (UK AISI, Epoch).
On Artificial Analysis's index, 60% based on secret tests since 27 October, 19 points separated the best closed model from the best open-weight model on 16 November; the gap was 26 points at the end of October.
What they don't say
What these 19 points represent in months; whether the internal acceleration measured by METR at Anthropic will last; how much revenue a month's lead brings.
A $70.7 billion bill for the coming year alone
The prospectus amended on 10 November puts compute and infrastructure commitments at $900.9 billion, of which $70.7 billion payable within twelve months after 30 September 2026 and $451.6 billion before the end of 2030. At the end of September, cash and investments totalled $131.8 billion: the coming year is covered. The offering, 825 million shares between $70 and $78, and a private placement with seven industrial companies should raise $84.6 billion net at the midpoint, for general corporate purposes, including compute and infrastructure. Against that, $35.3 billion in revenue over twelve months, including $17.0 billion in the third quarter, and an operating loss of $567 million in that quarter. Compute and infrastructure absorbed 58% of quarterly revenue.
The commitments are worth about 25 years of revenue at the pace of the last twelve months. That is the price of the compute constraint described in ActuIA's analysis of 21 October.
Four months' average lag, 19-point gap on a secret-test index
On 29 May 2026, Epoch AI calculated that, since January, the best open-weight models reached the level of closed ones with an average four-month lag, on a composite index. This lag was three months between 2023 and October 2025. The cyber domain gives similar figures. The UK AI Security Institute measured on 17 July a lag of four to seven months, against six to ten for open models released from January to September 2025. Epoch repeated the exercise on 16 October and finds six to seven months.
A public test has a flaw: its questions can end up in models' training data. Artificial Analysis changed its method on 27 October: 60% of its index now relies on kept-secret tests. Between the best closed model and the best published weights, the gap was 26 points in the new version, against 12 in the previous one, as ActuIA noted on 28 October. Artificial Analysis warns that scores from the two versions cannot be compared, and the index composition changed at the same time as the share of secret tests. The gap is now 19 points between Fable 5.2 and Kimi K3.1 Pro, 65 against 46. Artificial Analysis does not translate the points of this index into months. On its Progress Index, launched the same day, the body indicates that each record by open models since January 2025 has followed by two to eight months, four in median, the closed model that had reached that score; at the end of October, it placed Kimi K3.1 Pro about two months from the closed frontier, before the release of Fable 5.2.
Cost remains. For the same simulated attack path of 100 million tokens, the British institute puts the spend at about $85 with Opus 4.5 or 4.6, against an estimated $46 with GLM-5.2 and $1.19 with DeepSeek V4-Pro. In the Menlo Ventures survey, a fund that is an Anthropic shareholder, published on 3 November and conducted with nearly 400 US decision-makers, open source goes in one year from 11% to 19% of tokens consumed, but from 6% to 4% of spending. Anthropic goes from 40% to 55% of API spending by surveyed companies. Volume shifts to open models, money stays with frontier models.
What can widen the gap: models accelerating their own design
The METR report is the main public element supporting a growing lead, without itself measuring the gap with open models. Its evaluators examined the company, with extended access, from 12 August to 29 October. They conclude that internal model scores have been progressing since October 2025 1.9 times faster than before, with a confidence interval of 1.5 to 2.8. They assume, with high uncertainty, that a majority, perhaps 80%, of the current acceleration comes from the use of AI itself in research, and say they are confident it does not come from an increase in training compute. If a lab doubles its pace and its followers keep theirs, the gap in months lengthens.
METR's caveats weigh as much as the result. Including public models from OpenAI, Google and xAI, the acceleration falls to about 1.4 times, and statistical choices make it vary from 1.1 to 2.9. Anthropic redacted "the majority of the elements we judged most informative", on the grounds of not helping its competitors. METR neither requested nor obtained direct access to unpublished models, and the company withdrew during the mission the planned verification of compute allocation. The body judges it extremely unlikely that Anthropic sought to deceive it, but writes that the public must rely on its judgement far more than it should.
Claude Mythos 5.2, the same model as Fable 5.2 without certain protections, has been reserved since 9 November for verified organisations in cybersecurity and life sciences. Anthropic has also hardened its defences against distillation, the copying of a model by another trained on its responses. The prospectus finally shows a net revenue retention of about 900%: customers active a year earlier pay nine times more. The definition counts any customer, excluding Pro, Max and Ultra subscribers, having spent at least one dollar, in a reference quarter when the whole company billed $1.4 billion.
What makes it fragile: the prospectus says so itself
In its risk factors, Anthropic writes that open models could capture part of the market for certain uses and exert pressure on its prices, that differentiation between advanced models could shrink with comparable architectures and the rapid spread of research, and that competitors can sell at zero or negative margin. Its prices fell this autumn. Opus 5.5, released on 22 September, costs 20% less per token than Opus 5; Fable 5.2 also drops 20%; Haiku 5.5 is billed ten times less than Haiku 4.5 on requests under 100,000 tokens.
Prices fall, the compute bill does not. Some contracts, including Google's for at least $121.3 billion until 2033, are due regardless of usage. Only $186.2 billion of the $900.9 billion are cancellable or subject to conditions. Each price cut must therefore be offset by volume.
Research firm New Constructs draws the harshest conclusion. In a 6 October report relayed by CNBC, it argues that open models make profitability of closed models difficult, and writes: "we do not think Anthropic has a viable business". Anthropic did not respond to CNBC. Aswath Damodaran, finance professor at NYU, does not believe in this commoditisation in his 3 November post: winning companies retain customers to the point that they do not leave, even for "technically superior and cheaper" alternatives. He writes that Anthropic still has this work to do. The rampart he envisages is therefore not the model's lead, and it remains to be built. According to Reuters, investors met during the roadshow ask whether Anthropic will keep its lead against Google and its prices against comparable models; one of them says he did not get detailed answers, and the company did not comment.
The money raised does not buy a lead. It buys the right to buy it back at each model generation, which open models have so far caught up with four to seven months later, with fixed costs that do not fall when prices fall.
World models: others invest, no public work at Anthropic
A more radical threat would come from another family of systems. World models learn how the world reacts to an action, from images, videos or sensors rather than text, to act in the physical world. AMI Labs, co-founded by Yann LeCun, raised $1.03 billion in March on the thesis that language models will not suffice (on the researcher's positions, see ActuIA's article of 5 October). World Labs, by Fei-Fei Li, raised $1 billion in February, Reuters reported. Google DeepMind presents its Genie 4 as a step towards general AI and a training ground for robots and agents.
No public Anthropic work on world models was found by ActuIA in its research publications, in its prospectus, in its announcements or in the titles of its 670 open job offers on 16 November. The prospectus says the company focuses on software development and agent capabilities, and states it does not currently target video and image generation models. What exists extends language models. The Model Hardware Standard, presented on 27 August, connects agents to microscopes or machines via a common protocol, without training a world model. In Project Fetch, in June, Claude alone programmed access to a robot dog's sensors and detection of a ball, without managing to make it fetch, a progress that did not come from a targeted robotics effort, according to the team. On 9 November, Anthropic wrote that Fable 5.2 pilots robots markedly better, including on tasks for which it was never explicitly trained. Added to this are a chip design team, confirmed in August, and the Claude Science research workshop. According to The Information, Anthropic had preliminary discussions in the spring, which went no further, about acquiring Physical Intelligence, which develops robot models. The $900.9 billion is for chips and data centres, not for a model architecture. What a rupture would expose is the accumulated know-how on language models and the revenue curve that pays the bill.
For a CIO, the issue is commitment duration
A CIO building on Claude buys a supplier whose measured lead has so far been counted in months: open-weight models have caught up with closed models four to seven months after them, according to Epoch AI and the UK AISI. None of these measures covers Fable 5.2, and none says whether this pace will hold. The practical question is portability: being able to change models without rewriting one's agents, signing contracts whose duration does not exceed the gap they pay for, reserving the closed model for tasks where secret tests still show a wide gap.
The investor will read in the next filings the firm share of commitments, the weight of compute in revenue, 58% in the third quarter, and net retention on a broadened customer base.
Google DeepMind announced on 27 October the opening of Genie 4 to testers. Google is also an Anthropic shareholder and its compute provider; at the price to be set on 18 November, it must exercise an option to buy $10 billion of shares, provided for by a prior agreement.
Quotes translated from English by ActuIA. This article does not constitute investment advice.
Our articles will then appear first in Google Top Stories.
