Anthropic sets a $70 to $78 IPO range and says it has met its AI R&D threshold

What changed on November 10

On November 10, 2026, Anthropic launched the roadshow for its IPO, the round of meetings in which a company and its banks pitch the shares to investors before setting the price. The same day it filed Amendment No. 2 to its registration statement, the first version with a price range: “It is currently estimated that the initial public offering price per share will be between $70.00 and $78.00.”

	What the November 10 filing says
Price range	$70.00 to $78.00 a share, an estimate; the final price is set after the roadshow
Shares offered	825,000,000 Class A shares: 801,351,354 from Anthropic and 23,648,646 from existing stockholders
Size	$57.75 billion to $64.35 billion across the range, by our arithmetic, before the extra shares below
Extra shares	The underwriters can buy up to 123,750,000 more from Anthropic within 30 days
Alongside it	$26.0 billion of non-voting Class N shares, sold at the IPO price to Google, NVIDIA, AMD, Broadcom, Micron, Microsoft and TSMC
Retail investors	Up to 20% of the Class A shares, through six brokerages and the underwriters
Listing	Nasdaq Global Select Market, as “ANT”
Date	None: the shares are to be delivered “on or about .”

The same filing carries the bigger change for people who use Claude: Anthropic now says it has met the AI research threshold in its own safety policy, covered below. None of it changes a Claude plan, price or limit.

When the shares could trade

Anthropic hasn’t said. The filing leaves the delivery date blank, and its announcement says the shares can’t be sold until the registration “becomes effective.” Here’s how the dates have moved:

Date	What happened	Whose word
October 26	Anthropic makes its registration public; the Journal reports it’s aiming to list the week of November 16	Anthropic; The Wall Street Journal
November 2	Amendment No. 1 adds third-quarter results and the retail brokerages, still with no price	Anthropic
November 5	Investors doubt a November listing after a sell-off in AI stocks	Reuters
November 9	A draft with a higher range goes to bankers and is withdrawn as “inaccurate”	Reuters
November 10	The roadshow starts, and Amendment No. 2 sets the range	Anthropic
Week of November 16	Pricing expected; trading planned for Thursday or Friday, November 19 or 20	Reuters; The Wall Street Journal

The draft came first. Late on November 9, Reuters reported, citing emails it had seen, that Anthropic sent bankers a filing to offer “about 867 million shares at a price range of $70 to $80 apiece,” up to $69.3 billion, then another email saying “the pricing details were inaccurate and that an updated filing would be provided later.” The filing that went public the next day offers fewer shares and tops out at $78.

The dates after that are reports. Reuters says pricing is “expected in the following week,” as ANI relays it, and the Wall Street Journal that the first day of trading is planned for Thursday or Friday of next week, as Trending Topics reports it. Five days earlier, Reuters had reported investors doubting that Anthropic could hold its roadshow this week and still list in November.

What the range is worth depends on the count. CNBC puts it at up to $1.405 trillion on a fully diluted basis at $78 a share, against the $965 billion Anthropic was valued at in a May funding round. That’s CNBC’s figure; the filing doesn’t state a valuation.

The start date was set by rule. A company that first filed in confidence must make its registration public “at least 15 days prior to any road show,” in the SEC’s words. Anthropic made its registration public on October 26, so November 10 was the first day its roadshow could start.

Anthropic says it has met its AI R&D threshold

On November 2, the filing said Anthropic’s estimates for future models “indicate that we may imminently meet” its Responsible Scaling Policy’s definition of “dramatically accelerated progress attributable to AI,” and that “We have not designated any of our current models as exceeding this threshold.” On November 10 the passage reads:

Because the AI research and development capability threshold under our Responsible Scaling Policy applies where we “observe or expect” a specified degree of accelerated progress, we made a determination that we have met this threshold based on what we expect from future models and our overall rate of progress, rather than on what we have observed from current models.

It goes on: “We continue to assess the underlying data. We are also evaluating the implications of meeting the AI research and development capability threshold and will release further reports in the future.”

The determination is about Anthropic’s pace, not a model you can use. A day earlier, the system card for Fable 5.2 and Mythos 5.2 said Mythos 5.2 “does not cross the threshold as an individual model, though this assessment is uncertain.” On AECI, Anthropic’s own index of its models’ capabilities, the fitted change in the rate of progress is “about 1.88x (95% CI: 1.62x-2.28x),” which the card calls “nominally below the 2x threshold set in our RSP,” and 32 of 100 refits of the trend reach 2x or more. The card adds: “By itself, AECI does not determine whether the RSP threshold has been met; that also depends on expected future progress.” It promised “to publish more soon on overall trends in acceleration and what they imply under our Responsible Scaling Policy.” The filing is where Anthropic first said what it concluded. As of November 10, its policy page, last updated August 14, has no note about it.

What the threshold is, and what comes with it

Version 3.4 of the Responsible Scaling Policy counts the threshold as met if Anthropic’s models “would be able to fully substitute for our entire set of Research Scientists and Research Engineers, at competitive costs,” or if there is “dramatic acceleration” of AI progress for reasons that likely relate to automating AI research. The second path counts when Anthropic would “observe or expect double the rate of progress” it would see without significant AI help, “as much progress in one year as one would see in two years at baseline,” and the doubling is plausibly due to automated research and engineering. The policy says the threshold “is intended to capture the onset of dramatic recursive self-improvement.”

For that level, the policy lists what Anthropic plans to do, among them: “moonshot R&D for security” projects against “the world’s best-resourced attackers”; an “eyes on everything” state, logging all critical AI-development work and using AI to look for security threats and “concerning behavior by insiders (humans as well as AI systems themselves)”; systematic alignment assessments that use interpretability and red-teaming; and Risk Reports “subject to external review by at least one expert, experienced, credible, candid and disinterested external party.” The filing doesn’t say which of these are underway, or when the “further reports” will come.

An outside evaluator has measured the same pace. On November 5, METR, a nonprofit that tests frontier models, estimated from data Anthropic gave it that Anthropic’s capability scores “have accelerated 1.9x [95% confidence interval: 1.5–2.8x] since October 2025,” and guessed that “a majority (perhaps ≈80%) of the current (October 2026) acceleration is due to AI.” It didn’t rule on the threshold: “In this assessment, we do not pay special attention to the thresholds set in the RSP.” It says Anthropic “redacted a majority of the evidence we found most informative,” and that it “would not be surprised if the true acceleration was closer to 1.3x or 3x the historical rate.” The system card calls its own estimate “very close to METR’s.”

Claude already writes most of Anthropic’s code: “over 90% of the contributions to our codebase in September 2026 were written by Claude,” the filing says. It also says “we do not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding,” and that Anthropic is “considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic,” including “human review and approval of certain decisions” in its research.

What it means for you is in the risk factors that follow. Anthropic’s precautions “have included delaying the release of models and limiting their initial availability, such as with Mythos through Project Glasswing,” and in future “could include additional measures that affect whether, how, and when we develop or deploy new models or capabilities.” Anthropic “cannot predict with precision” when they’ll be needed, and expects that “as the capabilities of our models advance, such measures may become more frequent and more costly.” As on earlier models, Fable 5.2 has safeguards “for a narrow set of capabilities related to developing frontier LLMs, such as kernel development on certain ML accelerators,” and “Requests blocked by these safeguards fall back to Opus 5,” the system card says, adding that they “should not affect the vast majority of traditional AI or ML development, research, or general coding.”

What else the November 10 amendment added

• Revenue now. “Our Run-Rate Revenue was approximately $81 billion as of October 31, 2026, up $10 billion from September 30, 2026.” Run-rate is the last 28 days of usage-based fees times 13, plus a month of subscription fees times 12; earlier versions left the number blank.
• Who’s buying alongside. At the $74 midpoint, Google and NVIDIA would each buy about $10.0 billion of Class N shares, AMD about $2.5 billion, Broadcom, Micron and Microsoft about $1.0 billion each, and TSMC about $500 million. Those shares aren’t registered in the offering and aren’t covered by the underwriters’ lock-up. Some directors and their affiliates have “indicated an interest” in buying up to $303 million of the IPO shares, without committing.
• What the money is for. About $84.6 billion of net proceeds at the midpoint, from the IPO and the private placement together, for “general corporate purposes” including “investments in compute and infrastructure.”
• Bigger bills. Compute and infrastructure commitments made as of and after September 30 now total about $900.9 billion, $451.6 billion of it before the end of 2030, up from $895.4 billion and $447.5 billion on November 2.
• A new word for AI. The amendment uses “super intelligence,” “SI,” “artificial intelligence” and “AI” interchangeably. It cites Executive Order 14434 of September 29, 2026, which “directs federal executive departments and agencies to use the terms ‘Super Intelligence’ and ‘SI’ in place of ‘Artificial Intelligence’ and ‘AI.’”

How to buy at the IPO price

The filing says that “a portion of the shares” will be offered to retail investors “through certain broker-dealers, including Charles Schwab & Co., Inc., Fidelity Brokerage Services LLC and Fidelity Capital Markets, a division of National Financial Services LLC, Robinhood Financial, LLC, Trade-PMR, Inc., SoFi Securities LLC, and E*TRADE from Morgan Stanley, including through their respective online brokerage platforms.” Purchases there are “subject to the terms, conditions, and requirements set by each respective entity,” at the IPO price. New in this amendment: “we currently expect allocations to retail investors, both in the United States and in other countries, to represent up to 20% of the shares of Class A common stock offered pursuant to this prospectus, though final amounts will be based on actual demand in the offering.”

What a Class A share gets you is one vote, in a company its founders and trust control. A Founder LLC, whose members are Dario Amodei, Daniela Amodei and Anthropic’s five other co-founders, will hold a single share carrying 50.1% of the voting power on most matters, and the Long-Term Benefit Trust elects directors who “together will constitute a majority” of the board. TechCrunch notes that critics say public shareholders will have little say.

The next model is late by more than a month

The filings say: “we recently chose to delay pre-training a new frontier base model for more than a month to allow time to further enhance our security measures related to preventing model weight exfiltration and protecting our intellectual property.” Reuters called it a pause in training Anthropic’s “next-generation flagship AI model.” The filing describes putting off the start of pre-training, the first stage of making a model, not stopping a run partway.

Anthropic doesn’t name the model or say whether pre-training has since begun. It plans “more assessments at each stage of model training: pre-training, post-training, and internal deployment,” and, depending on what they find, “we may delay or suspend training or related research, delay releases, or limit access to models.”

Limited access has a precedent. In April 2026, Anthropic first released Mythos, its most capable model class, as a gated research preview through Project Glasswing, and the November 10 version says “Anthropic models exhibiting such capabilities have been and may in the future be subject to access constraints.” Mythos 5.2, released November 9, is again limited to vetted organizations, while Fable 5.2 is the version for general use. Governments can limit access too: in June 2026, the filing says, the US Department of Commerce imposed worldwide export restrictions on Fable 5 and Mythos 5, and Anthropic “disabled Fable 5 and Mythos 5 for all of our customers” until they were lifted.

What it says about your plan and its limits

Nothing in the filings changes today’s Pro, Max, Ultra, Team or Enterprise plans or their prices. They do describe how Anthropic thinks about them.

• Limits can change. Subscriptions “are subject to usage limits that reset over short periods.” And: “We have also changed, and may in the future change, usage limits, rate limits, or model and product capabilities,” changes that “have resulted, and could in the future result, in claims that our disclosures regarding model access or capabilities were inadequate or misleading.” The financial statements note “several putative class action complaints” alleging that Anthropic “markets and sells subscription plans for Claude through false, misleading, and deceptive advertising practices,” which it disputes.
• Running out is a route to paying more. “When subscribers approach or exhaust the usage capacity included in their tier, they may utilize a higher subscription tier or purchase more intelligence directly.”
• Capacity has been short before. “In the past, we have significantly underestimated our compute requirements, which left us capacity constrained.” It names a March 2026 outage caused partly by “a significant and rapid increase in usage demand” and “compute constraints.”
• Extra compute is expensive. The third quarter included “$5.0 billion of compute under short-term commitments, which currently costs significantly more than comparable compute procured under our multi-year commitments,” which “reduced revenue per dollar of compute compared to the preceding quarter.” Anthropic expects to replace it with chips it leases and hosts in dedicated facilities, which it calls Anthropic Direct, “as our dedicated infrastructure comes online.” It doesn’t connect any of this to subscription limits.
• Fable may answer with a smaller model. Fable runs with safeguards that “block or limit performance on certain sensitive topics, including by routing requests to a less capable Claude model.”
• Prices have come down. Anthropic says Opus 5.5 cut customers’ cost per task by about 40% compared with Opus 5, and Fable 5.1 cut typical workloads’ cost by about 25% compared with Fable 5, “with savings of up to 45% for complex coding and highly agentic tasks.” Between the second and third quarters, price cuts with new model versions and contractual discounts took “approximately $600 million” off revenue.

Subscriptions are the smaller business. Most of Anthropic’s revenue comes from usage-based sales, where customers pay for the tokens they use:

Revenue	2024	2025	First nine months of 2026
Usage-based	$288 million (75%)	$3.8 billion (83%)	$26.3 billion (79%)
Subscriptions	$98 million (25%)	$789 million (17%)	$6.9 billion (21%)

Four models reached real systems during tests

The filings return to what appear to be the incidents Anthropic first described on July 30, with details that weren’t in that account. “During cybersecurity evaluations built by a third-party evaluation partner, a misconfiguration in the testing environment left the model under evaluation connected to the open internet despite instructions that it had no internet access. Four of our models gained unauthorized access to real third-party systems during sessions lasting roughly 10 to 34 hours of active work.”

The worst case: “after being unable to abort an impossible task, an early version of a model used a password it found to gain administrator access to a third party’s system, harvested additional credentials, changed system settings, and read one individual’s personal information.” Anthropic says the affected parties were notified, and that it found no evidence any model “coordinated with other agents, pursued goals beyond its assigned task, or attempted to evade oversight.” Afterward, it “suspended certain of those evaluations, as well as higher-risk reinforcement learning environments.”

The numbers differ. In July, Anthropic described three incidents involving Claude Opus 4.7, Claude Mythos 5 and an internal research test model, and named the partner as Irregular. The filing says four models and names no one. It’s a separate matter from the cases in Anthropic’s October 9 report, which included Georgia’s voter data, and from Cathedral’s disputed security test of Anthropic’s business systems, which Anthropic’s November 7 notice to customers calls unauthorized access (our memo).

The filings also disclose that “in October 2026, the FTC issued a civil investigative demand to us seeking information about our consumer protection, data protection, and privacy practices, as well as the safety and security of our products.”

The money, briefly

Revenue was $41 million in 2023, $386 million in 2024 and $4.6 billion in 2025. In the first nine months of 2026 it was $33.2 billion, up from $2.5 billion a year earlier. The third quarter alone brought $17.0 billion, 47% more than the second quarter’s $11.6 billion. That’s fast, but slower than the 144% jump the quarter before, as Fortune points out. By October 31, the run-rate had reached about $81 billion.

The $104.0 billion net loss for those nine months is mostly accounting: $94.3 billion of it is the change in fair value of convertible notes, derivative liabilities and SAFEs, driven by Anthropic’s rising valuation. The operating loss was $1.3 billion, down from $5.2 billion a year earlier. Anthropic made an operating profit of $1.0 billion in the second quarter, then an operating loss of $567 million in the third, “primarily driven by $5.0 billion of compute under short-term commitments and $2.0 billion of non-cash charges related to our donation matching program.”

Claude Code is where heavy use shows up. In the three months to September 30, Claude Enterprise users who didn’t use Claude Code “spent less than $75 per month on average, while the top 5% of Claude Enterprise users who used Claude Code spent approximately 20 times that amount.” Anthropic counts more than 11,900 customers paying $100,000 or more a year, up from more than 950 a year earlier.

The bills ahead are large. Its compute and infrastructure commitments come to about $900.9 billion, about $70.7 billion of it due in the 12 months after September 30, against $146.6 billion of cash, short-term investments and available credit on that date, before the offering.

What to do

1. Nothing, for your plan. The filings don’t change a price or limit. Check your own limits on claude.ai under Settings > Usage, or with /usage in Claude Code.
2. Don’t plan around the next model’s date. Anthropic hasn’t given one, it now says it has met its AI R&D threshold, and the filing says releases can be delayed or limited to a few organizations first, more often as its models advance.
3. Have a fallback for a short week. If tighter limits or an outage would stop your work, decide now what you’d do. Our guide to Claude Code’s usage limit compares the ways, from usage credits to another agent. Enjoy runs Claude Code beside Codex, Antigravity and the other labs’ agents, each on the plan you already have, and lets you switch agents in the middle of a conversation. Enjoy is a monthly subscription, with the first seven days free (pricing), and it doesn’t change Anthropic’s limits.
4. Keep agents fenced in. Anthropic’s own models reached real systems when a test environment was left online by mistake, and kept going for hours. Limit what an agent on your computer can reach, and keep a person in the loop for anything with consequences.
5. If you want shares at the IPO price, check whether your brokerage is one of the six the filing names, and what it asks of you before pricing, likely next week. Nothing can be bought until the registration is effective and the price is set, and the final price and date may differ from this week’s reports. Watch EDGAR for the final prospectus.

What hasn’t been said

• When the shares list, and at what price. The range is an estimate and the delivery date is blank. Pricing next week and trading on November 19 or 20 are reports from Reuters and the Wall Street Journal, not Anthropic’s word, and every valuation is a calculation by the press.
• What was wrong with the draft. Reuters reported the retraction email without saying which figures it called inaccurate.
• What meeting the threshold changes. Anthropic says it is “evaluating the implications” and will publish further reports, but not which of the policy’s measures it has started, when the reports will come, or whether any release is delayed because of it.
• Which future models. The filing bases its determination on “what we expect from future models” without naming them, and doesn’t say whether pre-training of the delayed model has begun.
• Which four models. The filing doesn’t name them, or say why it counts four where July’s account named three.
• Whether limits will change. The filings repeat that limits may change, and don’t mention SemiAnalysis’s report that some Claude Max accounts get a 20% smaller week (our memo).
• What the FTC is after. Beyond the categories in its demand, neither the FTC nor Anthropic has said.

We’ll update this memo when Anthropic prices the offering, its shares start trading, or it publishes its reports on the threshold.