Wednesday, November 18, 2026

Top 5 This Week

Related Posts

Anthropic: 900 billion dollars for how many months in advance?


At the beginning of November, Moonshot AI put the weights of Kimi K3.1 Pro online, a model of 2800 billion parameters that any company can now download, within the limits of a license which governs commercial use. On the Artificial Analysis index, it obtains 46 points, the best score of the open weight models. On November 5, the evaluation organization METR published a report according to which Anthropic’s internal models have progressed approximately twice as fast since October 2025. Four days later, Anthropic released Claude Fable 5.2: 65 points. On Wednesday, November 18, the company is set to price its IPO, with $900.9 billion in computing and infrastructure commitments listed in its prospectus. Do these billions buy a lasting advance, or a few months?

What the measurements say

The best open models reach the level of closed ones four months late on average (Epoch AI, May 2026), four to seven months on cybersecurity tasks (UK AISI, Epoch).

On the Artificial Analysis index, based 60% on secret tests since October 27, 19 points separate the best closed model from the best open weight model on November 16; l’écart était de 26 points fin octobre.

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 income a month in advance brings in.

A bill of 70.7 billion for the coming year alone

The prospectus amended on November 10 figures computing and infrastructure commitments at $900.9 billion, including $70.7 billion to be paid in the twelve months following September 30, 2026 and $451.6 billion before the end of 2030. At the end of September, cash and investments reached $131.8 billion: the coming year is covered. The offering, 825 million shares between $70 and $78, and a private placement with seven manufacturers are expected to bring in $84.6 billion net at the median price, intended for general corporate needs, including computing and infrastructure. Opposite, 35.3 billion in revenue over twelve months, including 17.0 in the third quarter, and an operating loss of 567 million over this quarter. Compute and infrastructure absorbed 58% of revenue in the quarter.

The commitments are worth approximately 25 years of income based on the last twelve months. This is the price of the calculation constraint described in the ActuIA analysis of October 21.

Four months late on average, 19 points difference on a secret test index

On May 29, 2026, Epoch AI calculated that, since January, the best open weight models reached the closed weight level four months behind on average, on a composite index. This delay was three months between 2023 and October 2025. The cyber domain gives similar figures. The UK AI Security Institute measured a delay of four to seven months on July 17, compared to six to ten for the open models released from January to September 2025. Epoch repeated the exercise on October 16 and found six to seven months.

A public test has a flaw: its questions can end up in the model training data. Artificial Analysis changed its method on October 27: 60% of its index is now based on tests kept secret. Between the best closed model and the best published weights, the gap was 26 points in the new version, compared to 12 in the previous one, as noted by ActuIA on October 28. Artificial Analysis warns that the scores of two versions do not compare, and the composition of the index has changed along with the share of secret tests. The gap is today 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 organization indicates that each record of open models since January 2025 followed by two to eight months, four on median, the closed model which had reached this score; At the end of October, he placed Kimi K3.1 Pro about two months from the closed border, before the release of Fable 5.2.

There remains the cost. For the same simulated attack path of 100 million tokens, the British institute estimates the expense at around 85 dollars with Opus 4.5 or 4.6, compared to 46 dollars estimated with GLM-5.2 and 1.19 dollars with DeepSeek V4-Pro. In the survey by Menlo Ventures, Anthropic’s shareholder fund, published on November 3 and conducted among nearly 400 American decision-makers, open source increased in one year from 11% to 19% of tokens consumed, but from 6% to 4% of expenses. Anthropic increases from 40% to 55% of API spending of surveyed companies. Volume goes to open models, money stays to frontier models.

What can widen the gap: models that accelerate their own design

The METR report is the main public element supporting a growing lead, without itself measuring the gap with open models. Its reviewers reviewed the company, with extensive access, from August 12 to October 29. They conclude that the scores of the internal models have progressed 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 is due to the use of AI itself in research, and say they are confident that it does not come from an increase in training computing. If a laboratory doubles its pace and its followers keep theirs, the gap in months lengthens.

METR’s reserves weigh as much as the result. Including the public models of OpenAI, Google and xAI, the speedup drops to around 1.4 times, and statistical choices vary it from 1.1 to 2.9. Anthropic redacted “the majority of the elements that we considered to be most informative,” on the grounds of not helping its competitors. METR neither requested nor obtained direct access to the unpublished models, and the company withdrew the planned verification of the calculation allocation during the mission. The organization considers it extremely unlikely that Anthropic sought to deceive it, but writes that the public must rely on its judgment much more than it should.

Claude Mythos 5.2, the same model as Fable 5.2 without certain protections, has been reserved since November 9 for verified organizations in cybersecurity and life sciences. Anthropic has also hardened its defenses against distillation, this copying of one model by another trained on its responses. The prospectus finally shows a net revenue retention of around 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 where the entire company billed 1.4 billion.

What makes it fragile: the prospectus writes it 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 the differentiation between advanced models could be reduced with comparable architectures and the rapid diffusion of research, and that competitors can sell at zero or negative margin. Ses tarifs ont baissé cet automne. Opus 5.5, released on September 22, costs 20% less per token than Opus 5; Fable 5.2 baisse aussi de 20%; Haiku 5.5 is charged ten times less than Haiku 4.5 on requests of less than 100,000 tokens.

Prices are falling, the calculation bill is not. Some contracts, including Google’s for at least 121.3 billion until 2033, are due regardless of use. Only 186.2 billion of the 900.9 billion are cancellable or subject to conditions. Each price reduction must therefore be compensated by volume.

The research firm New Constructs draws the harshest conclusion. In an October 6 report relayed by CNBC, he argues that open models make it difficult for closed models to be profitable, writing: “we do not believe Anthropic has a viable business.” Anthropic n’a pas répondu à CNBC. Aswath Damodaran, professor of finance at NYU, does not believe in this trivialization in his post of November 3: companies that win retain their customers to the point that they do not leave, even towards “technically superior and cheaper” alternatives. He writes that Anthropic still has this work to do. The rampart he envisages is therefore not the advance of the model, and it remains to be built. According to Reuters, investors met during the roadshow asked whether Anthropic will maintain its lead against Google and its prices against comparable models; one said he did not get detailed answers, and the company had no comment.

Money raised does not buy an advance. He buys the right to buy it back with each generation of models, which the open models have so far joined four to seven months later, with fixed costs that do not fall when prices fall.

World models: others invest in them, 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 be enough (on the researcher’s positions, see the ActuIA article of October 5). Fei-Fei Li’s World Labs 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 work from Anthropic on world models was found by ActuIA in its research publications, in its prospectus, in its announcements or in the titles of its 670 job offers opened on November 16. The prospectus says the company is focused on software development and agent capabilities, and says it is not currently targeting video and image generation models. Ce qui existe prolonge les modèles de langage. The Model Hardware Standard, presented on August 27, connects agents to microscopes or machines through a common protocol, without training a model of the world. In Project Fetch, in June, Claude alone programmed access to the sensors of a robot dog and the detection of a ball, without succeeding in having it brought back to him, a progress which did not come from a targeted effort on robotics, according to the team. On November 9, Anthropic wrote that Fable 5.2 drives robots significantly better, including tasks for which it was never explicitly trained. Added to this is 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 remained without follow-up, with a view to buying Physical Intelligence, which develops models for robots. The $900.9 billion is for chips and data centers, not model architecture. What a breakup would expose is the accumulated know-how about language models and the revenue curve that pays the bill.

For a CIO, the issue is the duration of commitment

A CIO who builds on Claude buys a supplier whose measured advance has so far been counted in months: open weight models reached the level of closed models four to seven months after them, according to Epoch AI and the British AISI. None of these metrics relate to Fable 5.2, and none say whether this pace will hold. The practical question is portability: being able to change models without rewriting your agents, signing contracts whose duration does not exceed the gap they pay, 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 the calculation in revenues, 58% in the third quarter, and the net retention on an expanded customer base.

Google DeepMind announced on October 27 the opening of Genie 4 to testers. Google is also a shareholder of Anthropic and its computing supplier; at the price to be set on November 18, he must exercise an option to purchase $10 billion in shares, provided for in a previous agreement.

Quotes translated from English by ActuIA. This article does not constitute investment advice.

Our articles will be prioritized in Google’s “Featured” section.

Shredder Smith
Shredder Smith
Shredder Smith is the lead curator and digital persona behind topaitools4you.com, an AI directory dedicated to "shredding" through industry hype to identify high-utility software for everyday users. Smith positions himself as a blunt, no-nonsense reviewer who vets thousands of emerging applications to filter out overpriced "wrappers" in favor of tools that offer genuine ROI and practical productivity. The site serves as a watchdog for the AI gold rush, providing categorized rankings and transparent reviews designed to help small businesses and creators navigate the crowded tech landscape without wasting money on low-value tools.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Popular Articles