Anthropic's Amended S-1 Says It May Imminently Meet Its Own Threshold for AI-Accelerated Progress. The Same Filing Shows a $567 Million Third-Quarter Operating Loss.
On Monday, November 2, 2026, Anthropic filed Amendment No. 1 to the S-1 it made public on Monday, October 26, seven days earlier. Fortune, TNW and others led with the third-quarter numbers the first filing left blank. The sentence I keep rereading is in the risk factors.
Anthropic writes that its internal estimates and its expectations about future models, "rather than our current models, indicate that we may imminently meet" its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI." TechCrunch's headline put it as Anthropic being "very close to RSI capability," meaning recursive self-improvement. That is a fair shorthand with one missing qualifier, which the filing supplies in the next breath.
What the Risk Factor Says
I could not pull the full filing myself; at more than 10 MB, the SEC document was too large for our fetch tools. The passages below are those quoted in matching words by at least two of the outlets linked at the end, plus TechCrunch's summary.
- Not there yet: "We have not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data."
- Late detection: Anthropic says "we have in the past not identified, and may in the future not identify, a significant acceleration in our rate of development" until after it has happened, and it says its evidence on AI-driven speedups comes from lagging indicators.
- The safety risk: "Using AI to help build more capable AI could also pose a considerable safety risk," particularly if it sharply accelerates progress. The filing says this is covered by its Responsible Scaling Policy, which sets out plans to monitor and mitigate the risks "as we approach what has been termed 'recursive self-improvement'".
- Human oversight: Anthropic says it does not expect to safely let Claude fully autonomously design and build successive generations of frontier models without meaningful human supervision, and it is considering RSP updates to "keep humans in control of frontier model development."
Two comparisons we read, by Cellcog and Data Studios, say the "may imminently meet" language is new in the amendment and was not in the October 26 filing. We have not checked that against the original S-1 line by line.
What the Threshold Actually Is
The words "dramatically accelerated progress" carry a specific meaning in Anthropic's own policy. The current RSP is Version 3.4, effective July 8, 2026. Its AI R&D threshold has two routes, and the filing's language points at the second.
| Route | RSP 3.4 wording, summarized |
|---|---|
| (1) Substitution | Models could fully substitute for all of Anthropic's Research Scientists and Research Engineers at competitive costs (within a factor of 5) |
| (2) Acceleration | Anthropic observes or expects double the rate of progress in aggregate AI capabilities, against both the expected rate and the fastest extended rate seen without major AI help |
| Attribution | It must be plausible that the doubling is substantially attributable to automating research or engineering |
| "Double" | "as much progress in one year as one would see in two years at baseline"; a 9x baseline effective scaleup would become something like 81x |
| Purpose | "intended to capture the onset of dramatic recursive self-improvement, and has proven difficult to operationalize" |
Note the word "expects." The threshold can be met by forecast, not only by measurement. That is why a sentence about future models matters: by Anthropic's own rules, an internal expectation of a doubling is enough to count.
The Quarter Next to It
The financial half of the amendment is less abstract. In October we wrote that the second quarter was Anthropic's only profitable quarter of the six the S-1 reported. The third quarter did not extend the streak.
| Line | Q2 2026 | Q3 2026 |
|---|---|---|
| Revenue | $11.56 billion | $16.96 billion (up about 47%) |
| Total operating expenses | $10.6 billion | $17.5 billion |
| Operating result | $974 million profit (8.4%) | $567 million loss (minus 3.3%) |
| Compute and infrastructure | $5.56 billion (48% of revenue) | $9.86 billion (58% of revenue) |
Margins are our arithmetic. Line items come from the filing as reported by Fortune, PPC Land and TNW; the compute rows are PPC Land's reading of the filing.
Anthropic attributes the loss primarily to "$5.0 billion of compute under short-term commitments, which currently costs significantly more than comparable compute procured under our multi-year commitments," plus about $2.0 billion of non-cash charges tied to its employee donation matching program. Fortune adds that the filing cites a decrease of approximately $600 million from price reductions between the two quarters.
The $23.3 billion third-quarter net loss is mostly an accounting charge on convertible notes whose value rose, per Fortune and TNW, and is not a cash figure. The commitments are real: about $895.4 billion of purchase commitments, substantially all for compute and infrastructure. Liquidity at September 30 was $146.6 billion including available credit line capacity. Revenue for the first nine months of 2026 was $33.2 billion.
Our Take
TechCrunch noted skeptics who argue an S-1 is partly marketing. I do not buy that framing in full. A risk factor is a legal disclosure, not a press release, and this one includes an admission that does not sell anything: Anthropic says it has missed accelerations before and may miss them again. That is the opposite of a capability boast.
But nothing in it can be checked from the outside. The threshold can be met by expectation, the data is internal, the indicators lag, and the filing says no current model is over the line. Investors are being asked to price a forecast whose inputs they cannot see. Anthropic has published its own R&D automation index before, which we covered in September; a crossing call of this weight deserves the underlying series, not a sentence.
The two halves of the filing also pull against each other in an interesting way. One half says AI may soon double Anthropic's rate of progress. The other says the company paid a premium for $5.0 billion of short-term compute in a single quarter. If automated research is close to compounding, compute is the input it compounds on, and Anthropic is still buying some of it at short-term rates it calls significantly more expensive. That is the operating number to watch more than the net loss.
Three Signposts for the Next 60 Days
First, whether Anthropic publishes the RSP update it says it is considering to keep humans in control of frontier development, and what it actually requires. Second, whether Anthropic formally designates any model as meeting the AI R&D threshold, before or after the listing. Third, whether short-term compute falls back as a share of costs in the fourth quarter as long-term capacity comes online.
Sources: Anthropic S-1/A, SEC, November 2, 2026, Anthropic Responsible Scaling Policy (Version 3.4 PDF), TechCrunch, Fortune, PPC Land, TNW, Data Studios, Cellcog, Dataconomy.