Anthropic Revises Its Prospectus, Admits Claude Can’t Replace Its Own R&D Team

Anthropic Revised Its Prospectus, Admits Claude Cannot Replace Its Own R&D Team

Just now, Anthropic revised its IPO prospectus.

Compared with the original version on October 26, a new sentence was added to the risk disclosure:

Currently, no model has crossed the RSP threshold for AI research and development.

Simply put, it’s about whether AI can turn around and help us accelerate the development speed of new models.

Previously, Anthropic publicly stated that Claude contributed over 90% of the company’s code and is increasingly involved in the development of next‑generation models.

The story of AI helping AI evolve has been heard many times.

But in the documents formally submitted to regulators, Anthropic gave this assessment…

It sounds like they’re pouring cold water on themselves.

◈1. Why hasn’t the threshold for accelerating R&D been crossed yet?

RSP stands for Responsible Scaling Policy.

This is an internal policy Anthropic uses to assess the risks of frontier models. As a model’s capabilities increase, stricter safety measures must be applied.

The thresholds cover several areas such as chemical and biological weapons, and one of them is AI research, which looks at whether AI can significantly accelerate Anthropic’s own model development.

There are two trigger mechanisms; meeting either one counts as crossing the line.

First, AI can replace an entire R&D team.All of Anthropic’s research scientists and engineers could be substituted by AI, and the cost would not exceed five times the expense of maintaining that team.

The second type, AI makes the entire industry’s progress noticeably faster.Fast enough to cover two years’ worth of progress in just one year.

First, look at the first type.

Anthropic has an internal test called CoBench, which puts the model back into the company’s codebase from a past point in time, letting it investigate problems that engineers actually solved back then.

According to Anthropic’s estimate, a model that could truly replace a development team would need to achieve at least 85% on this test. But the best result in the chart is 62.8%, still far short.

In other words, even the best model inside Company A,AI still cannot replace a development team.

Now look at the second scenario.

The second scenario focuses on speed—whether Anthropic’s own models have improved overall capability at twice the rate of the past.

How do we measure speed?

Anthropic uses a primary tool called AECI, which aggregates a large number of internal test scores into a single metric, assigning each model a point.

It looks like this:

The dashed line in the image is the trend line connecting the eight frontier models up to Claude Opus 4.6, averaging an increase of 13.5 points per year. If each new model pulls further ahead of this line, it indicates acceleration.

In practice, Claude Mythos Preview indeed jumped above the line.

Anthropic asked the researchers involved in the development, and they believe this leap stems from several specific research breakthroughs, which themselves were not largely driven by AI.

Therefore, the report concludes that Anthropic’s R&D indeed accelerated noticeably starting in the first half of 2025, but the increase is less than double.

This round of acceleration, Anthropic is fairly confident, is mainly driven by non‑AI factors.

Meanwhile, Company A is considering revising the RSP to require certain R&D decisions to undergo human review.

In the end, it boils down to returning to basics…

◈2. Q3 Financial Data

The revised edition also completes the Q3 financial data.

Q3 revenue was $16.956 billion, a month-over-month increase of about 47%.

However, after recording an operating profit of $974 million in Q2, Q3 turned into an operating loss of $567 million.

Revenue grew, but profit didn’t hold.

The main reason is that compute costs on the expense side have risen too quickly.

Company A reported that in Q3, compute and infrastructure expenses reached $9.863 billion, up about $4.3 billion from the previous quarter. Of that, $5 billion came from higher‑priced short‑term compute.

◈3. 7% of revenue comes from non‑compliant accounts

When it comes to revenue, there’s also a figure that directly relates to users.

Anthropic previously disclosed that in the first nine months of this year,about 5% to 7% of revenue came from customers later found to be violating usage policies or commercial terms.

Company A also reminds investors that, as detection and enforcement measures intensify, the proportion of this revenue is expected to decline.

Recently, Company A updated the Claude usage policy, which will take effect on November 12.

The following actions are explicitly prohibited:

1. Violating the supported region policy.
2. Creating a new account after being banned or borrowing someone else’s account to continue using the service will result in a ban. 【Collective responsibility】

Sigh… After finally integrating Claude into the workflow, the process just started running smoothly.

AI self‑evolution is still underway, and the account suspension email may arrive first.