SafetyRisk, alignment & guardrails
Anthropic tells SEC it has met its threshold for AI-accelerated AI research ahead of its $83-per-share IPO
Anthropic's amended IPO prospectus says the company has met its Responsible Scaling Policy threshold for AI speeding up AI research. It bases this on expected future models rather than a current one. Separately, METR estimated internal progress is running about 1.9 times faster.
From the Financial Times · AI article of November 19, 2026 and cited sources: The prospectus language shifted over time. [2] [6] [7] [8]
SOURCES[2] CASRAI: Anthropic's IPO filing now says it has met its own… casrai.org[6] The Gadgeteer: AI-builds-AI threshold may be imminent the-gadgeteer.com[7] AI Weekly: Anthropic tells SEC its self-improvement… aiweekly.co
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Financial Times · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 19, 2026 article published
- Covered
- Nov 19 edition
- Evidence
- 11 sources across 11 sites
- 1AI advances at Anthropic trigger new safety fears on eve…ft.com
- 2CASRAI: Anthropic's IPO filing now says it has met its own…casrai.org
- 3METR: Independent assessment of AI R&D acceleration at…metr.org
- 4Yahoo Finance: Anthropic IPO livefinance.yahoo.com
- 5ABC7/CNN: Anthropic set to make stock market debutabc7news.com
- 6The Gadgeteer: AI-builds-AI threshold may be imminentthe-gadgeteer.com
Reach: not measured yet
The prospectus language shifted over time. The original S-1 did not mention the threshold. A Nov 2 amendment said Anthropic might soon meet it, and a Nov 10 amendment said it had met it. Under the policy, the threshold means compressing two years of progress at the 2018–2024 rate into one year. No current model has been named as crossing it. [2] [6] [7] [8] [9]
METR spent about eleven weeks working inside Anthropic with high access. It estimated a 1.9x acceleration (95% CI 1.5–2.8x) on an internal capabilities index and attributed roughly 80% of it to AI. METR did not judge whether the policy threshold was met. Its independent estimate rests on company data, and much of that data was redacted. [3]
The filing makes safety policy a disclosed investment risk: Anthropic says future safeguards could shift when it develops and releases models. The 1.9x speed-up is an estimate with wide uncertainty.
Read the full assessment
Staggered releases and new obligations are possible consequences that have not happened yet.
Executive brief
Anthropic told the SEC in its IPO prospectus that it has met its own threshold for "dramatically accelerated progress attributable to AI", the point where AI speeds up AI research itself. An outside evaluator, METR, measured internal capability growth at roughly 1.9x its earlier rate. The FT (FT) reports that the company disclosed it had activated part of its Responsible Scaling Policy (RSP).
Read the full section
Anthropic told the SEC in its IPO prospectus that it has met its own threshold for "dramatically accelerated progress attributable to AI", the point where AI speeds up AI research itself. It made that call based on what it expects from future models, not on a capability it has already seen in a current one (CASRAI). An outside evaluator, METR, measured internal capability growth at roughly 1.9x its earlier rate. METR also said Anthropic redacted most of the evidence it found most informative (METR). The FT (FT) reports that the company disclosed it had activated part of its Responsible Scaling Policy (RSP). That happened on the eve of a listing priced at $83 per share (Yahoo Finance).
What changed and event timeline
Original S-1 filed
The first prospectus said Anthropic might not spot a big acceleration in its own development until after it had happened. It did not mention the threshold (;).
Amendment 1 says "may imminently meet"
Anthropic's internal estimates now pointed to the RSP acceleration threshold being close. The filing said no current model had been designated as exceeding it ().
METR publishes an independent assessment
After high-access work inside Anthropic from Aug 12 to Oct 29, METR estimated 1.9x acceleration on an internal benchmark. It did not assess the RSP threshold itself ().
Amendment 2 says "has met"
Anthropic stated it has met the threshold, based on expected future models and its overall rate of progress (;).
19: IPO priced, then trading
Shares priced at $83, raising a reported $81.3B. Trading on Nasdaq under the ticker ANT began Nov 19 ().
Capabilities and access
- No model has been named as crossing the threshold. Amendment 1 said no current model had been designated (The Gadgeteer).
- TNW reports that a model called Claude Fable 5.2 was released Nov 9, and that its system card said it does not double progress.
- The filing cites earlier precautions such as delayed releases and limited initial availability (CASRAI).
Read the full section
- No model has been named as crossing the threshold. Amendment 1 said no current model had been designated (The Gadgeteer).
- TNW reports that a model called Claude Fable 5.2 was released Nov 9, and that its system card said it does not double progress. That name and claim are not confirmed elsewhere in the reviewed sources (TNW).
- The filing cites earlier precautions such as delayed releases and limited initial availability (CASRAI).
Technical analysis for researchers and developers
- What the threshold means: RSP v3.4 (effective July 8, 2026) defines it as AI that can compress two years of 2018–2024-rate progress into one year.
- How METR measured it (METR):
- Main metric: Anthropic's internal version of the Epoch Capabilities Index ("AECI").
Read the full section
- What the threshold means: RSP v3.4 (effective July 8, 2026) defines it as AI that can compress two years of 2018–2024-rate progress into one year. It is about doubling the pace of aggregate capability gains, not individual researcher productivity (Anthropic RSP).
- How METR measured it (METR):
- Main metric: Anthropic's internal version of the Epoch Capabilities Index ("AECI").
- Method: regression with a detected trend break on Oct 7, 2025.
- Other inputs: compute and headcount data, about 15 staff interviews, and self-reported uplift polls. METR treats the 12x figure from those polls as loose evidence only.
- Reproducibility: low. The work ran on company devices, and most key evidence was redacted.
Claims and evidence
- Threshold met (Nov 10)
- 1.9x acceleration (95% CI 1.5–2.8x)
- About 80% of the acceleration comes from AI
Read the full section
| Claim | Basis |
| Threshold met (Nov 10) | Company judgment based on expected future models; underlying data not published (AI Weekly) |
| 1.9x acceleration (95% CI 1.5–2.8x) | Independent METR estimate, but it relies on company-supplied data (METR) |
| About 80% of the acceleration comes from AI | METR estimate, with substantial uncertainty stated (METR) |
| Internal Claude usage value grew more than 250x from Jan to Sep 2026 | Filing figure, as reported by The Gadgeteer |
Context and prior work
- Anthropic first activated ASL-3 protections with Claude Opus 4 in 2025.
- In the policy's own words, it was becoming "increasingly difficult" to confidently rule out the AI R&D threshold.
- The prospectus has 93 pages of risk factors, against roughly 30 in Snap's (ABC7/CNN).
Read the full section
- Anthropic first activated ASL-3 protections with Claude Opus 4 in 2025. These included classifiers to block bioweapons misuse and stronger security against model-weight theft (TIME).
- In the policy's own words, it was becoming "increasingly difficult" to confidently rule out the AI R&D threshold. Anthropic had already committed to publishing sabotage risk reports for frontier models well beyond Claude Opus 4.5 (Anthropic RSP).
- The prospectus has 93 pages of risk factors, against roughly 30 in Snap's (ABC7/CNN).
Limitations, safety and contested findings
- METR's caveats (METR):
- The estimate could be too high: non-AI factors, gains limited to benchmark tasks, or an unusually slow historical baseline.
- It could be too low: missed recent speed-ups, or autonomous AI contributions arriving sooner than expected.
Read the full section
- METR's caveats (METR):
- The estimate could be too high: non-AI factors, gains limited to benchmark tasks, or an unusually slow historical baseline.
- It could be too low: missed recent speed-ups, or autonomous AI contributions arriving sooner than expected.
- Vague obligations: the safeguards the threshold triggers ("unprecedented levels of security", "eyes on everything", red-teaming itself) come with no published metrics (AI Weekly).
- Critics:
- Two researchers on Bluesky called the disclosure IPO-timed spin (AI Weekly).
- The AI Futures Project called for an industry-wide pause and warned that human-review safeguards "could provide too much comfort" (TNW).
Business and practitioner implications
- Investors: the RSP is now a disclosed risk factor. Anthropic says future safety measures could change the timing of development and releases (CASRAI).
- Builders: expect staggered or limited model releases. Anthropic says it does not expect Claude to safely build its successors without meaningful human supervision (Yahoo Finance).
- Evaluators: METR's high-access, partly redacted review sets a precedent for audits of this kind.
Read the full section
- Investors: the RSP is now a disclosed risk factor. Anthropic says future safety measures could change the timing of development and releases (CASRAI). Columbia's John Coffee had earlier flagged securities-liability exposure from safety statements (The Daily Upside).
- Builders: expect staggered or limited model releases. Anthropic says it does not expect Claude to safely build its successors without meaningful human supervision (Yahoo Finance).
- Evaluators: METR's high-access, partly redacted review sets a precedent for audits of this kind.
The source trail.
Sources (11)
AI advances at Anthropic trigger new safety fears on eve of IPO
Feed excerpt only.
ft.com