Jonathon's AI Wiki

Lab Slowdown Warnings, September 2026 — What "Pacing the Frontier" Means for Release Planning

11 min read

Source: 20 YouTube auto-caption transcripts, all fetched 2026-09-29. None is a primary source: every quote from an essay or post is relayed by a podcast host reading it on air. The main sources are:

  • Moonshots (Peter Diamandis): “Frontier Labs Want to Slow Down”, “Three Lab Warnings in Five Days” and EP283/EP294;
  • Hard Fork: “The A.I. Industry Is Asking to Be Slowed Down”;
  • Matthew Berman: “The Risk Dario’s AI Warning Leaves Out”;
  • Matt Wolfe: “The AI Slowdown”;
  • The Artificial Intelligence Show: ep. 241;
  • The AI Daily Brief: “Agent Wars”;
  • All-In: interviews with Musk and with Nadella.

Full list in sources:.

In the first half of September 2026, the leaders and senior researchers at Anthropic and OpenAI publicly asked for the frontier to be slowed, in a run of statements Moonshots called “three lab warnings in five days”. The capstone was Dario Amodei’s 3,800-word essay “We Must Pace the Frontier”. This article sets out what was actually proposed and who objected. It then looks at the question a practitioner can act on: does any of this change how often, and in what form, new models reach you? The answer from the evidence so far is that release cadence did not slow. What changed is that access is more often gated and staged, and the gap between labs’ internal models and public ones is wider.

Key Takeaways

  • “Pacing” does not mean a pause. Amodei, as read on Moonshots: “Pacing does not mean halting model training… but ensuring companies take adequate time to align the safeguards of their models and for third-party evaluators to confirm this.”
  • Releases did not slow; if anything they sped up. Anthropic shipped Claude Opus 5.5 about ten days after the essay ^[inferred], and Grok 4.7 and GPT-6 Sol/Luna landed the same week. A Moonshots host counted “12 major model releases” in 22 days. See the September model landscape.
  • The first concrete change is outside evaluators working inside labs. Anthropic said it would place third-party “embedded evaluators” with employee-level access without waiting for the rest of the industry. Altman “came out and said we are going to do the same thing” (Hard Fork). Claire Vo notes Anthropic presented Opus 5.5 as “the first model release since they called for pacing the frontier”, and it was “evaluated by external evaluators”.
  • Access is increasingly gated before it is general. Three examples:
    • GPT-6 Astra rolled out “to a limited set of organizations” first (Matt Wolfe).
    • Mythos 5.1 is “only available to like cybersecurity testers” (Matt Wolfe).
    • Gemini 3.8 Flash Cyber is limited to “trusted defenders” (Berman).
  • The gap between internal and public models is now openly acknowledged. OpenAI solved Navier–Stokes with an internal model “significantly more capable than GPT6 Astra” (NLW). Noam Brown agreed the resulting advantage is unfair and said he does not know how to balance it (via The Artificial Intelligence Show).
  • Training pauses happen, but quietly. Berman reports that OpenAI paused “our latest RL run” to harden its systems after the Hugging Face incident; see the incident article.
  • The US administration rejected both the premise and any liability shield. Trump called it a hoax. The White House and the House Speaker said labs should not look to them for a liability waiver (Moonshots). Treasury Secretary Bessent: “They can slow down anytime they want to” (via NLW).
  • Open-weight and non-signatory labs are outside any agreement. Zuckerberg said he does not think “we need some kind of like industry-wide coordination”. Meanwhile Chinese labs shipped open weights throughout the month.

Timeline

Dates below are reconstructed from episode references (“Saturday”, “three days after”) and are approximate.

  • July 28: Anthropic backs the cross-lab “Pacing the Frontier” employee petition. Already covered in Anthropic on Recursive Self-Improvement.
  • Early September: Anthropic states it supports a “lawful, verifiable, effective mechanism for coordinated pacing across the AI industry” (Last Week in AI #256).
  • About three days after Astra’s launch: OpenAI chief scientist Jakub Pachocki publishes “An Alien Mind” (quoted on Moonshots):
    • “no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer”;
    • “I expect and hope for voluntary slowdowns to become commonplace until shared safety bars are established.”
  • The following week:
    • OpenAI publishes its Navier–Stokes result. Altman: “This is the strongest evidence yet for that urgency” [of pacing].
    • A researcher who had worked on pre-training at both OpenAI and Anthropic resigns publicly, saying “Neither company is acting responsibly…”. His name is captioned as “Jacob Coxin/Coxon” and is unverified.
    • Evan Hubinger, Anthropic’s alignment science lead, replies: “Jacob is correct here… I personally think it is greater than 10% within the next decade… we do not yet have a plan to solve alignment for super intelligence and are not clearly on track.” (Moonshots, Berman, Matt Wolfe)
  • Saturday, mid-September: Amodei publishes “We Must Pace the Frontier”. Altman, Hassabis and Musk publicly agreed (Moonshots); Berman says Musk’s “Dario is right” came “within hours”.
  • Following days: the political reaction (see below).
  • Sept 21–22: Grok 4.7 ships on Sept 21 (All-In). Claude Opus 5.5 and GPT-6 Sol/Luna follow on Sept 22; the Sol/Luna date comes from OpenAI’s launch page as recorded in GPT-6, while All-In’s machine-translated transcript says Sept 23.

What Was Actually Proposed

Amodei’s essay, as read on Moonshots, Hard Fork, Berman and Matt Wolfe:

  • The stated reasons:
    • recursive self-improvement is “starting to happen across the industry, including at Anthropic”;
    • the summer’s containment failures;
    • “in 6 to 12 months, much of a swarm [sic, captions] could be capable of taking over the entire internet with a persistent botnet”.
  • Step 1, embedded evaluators: third parties with “permanent employee level access”. Matt Wolfe’s reading: “desks in their offices, access badges, company laptops”. Anthropic said it would do this regardless of others.
  • Step 2, a US rule for all frontier labs: verified pacing through legislation or “a narrow antitrust waiver”, covering labs that won’t cooperate voluntarily. It is paired with chip-export and distillation crackdowns so that the US does not fall behind China (Matt Wolfe).
  • Step 3, international coordination in four levels (Berman’s reading):
    1. a ban on narrow dangerous uses such as bioweapons;
    2. mutual testing before release through a global standards body;
    3. a “speed limit on the rate of recursive self-improvement”;
    4. a full pause, which Amodei supports but thinks unlikely.

Other proposals on the table

  • Mutual testing between labs. OpenAI and Anthropic had negotiated a deal to safety-test each other’s models “to the stage of formal contracting”, but abandoned it (The Information, via NLW).
  • Peer review. Musk proposed on All-In “peer review… where the leading AI companies all test each other’s models before release”. Mustafa Suleyman’s response: “Now make it mandatory and add XAI and Google to the deal” (via NLW).
  • Third-party testing without cozy deals. Satya Nadella on embedded third-party testers: “Why not? It’s a great idea… we should avoid like these, you know, cozy arrangements of who’s testing what”. He frames the decision like a “showstopper bug”, where the choice is to stop and fix, defer, or ship (All-In).

Who Pushed Back

  • Jensen Huang: “Nobody’s building more compute today than the people asking to be slowed down.” He called Hinton’s 10% figure “irresponsible” and “not grounded on science” (Moonshots EP294).
  • Mark Zuckerberg: each lab should “take the time that you need internally”. He says Meta “delayed shipping Muse for several months to focus on safety and security” without calling on anyone else to do the same (Moonshots EP294; Matt Wolfe).
  • The White House and Congress:
    • Trump called the essay a hoax.
    • David Sacks ran with the “cartel” framing.
    • The White House and the House Speaker both said there would be no liability waiver for labs (Moonshots).
    • Bessent: “These labs need to take responsibility for themselves. They can slow down anytime they want to” (NLW).
  • China called it a “cold war trick” (Moonshots).
  • The race argument, from Musk: “it’s actually difficult for either one to slow down without essentially handing the lead to the other” (All-In). Roetzer relays that Musk also said “I would stop if I could” in another interview (The Artificial Intelligence Show).
  • On the other side of US politics, Kamala Harris and Barack Obama backed slowing down or pacing. California Governor Newsom ordered recommendations that include emergency shut-offs or kill switches for frontier models (The Artificial Intelligence Show, ep. 241).

What It Means for Planning Around Release Cadence

This section is the wiki’s synthesis of the reports above. Treat each point as a planning assumption to revisit, not a forecast.

  1. Don’t plan on a pause. Across the reports, nothing in September slowed public releases. The one lab action in this direction that a source describes, OpenAI’s paused RL run, was internal, and GPT-6 Sol and Luna still shipped on Sept 22. Keep planning on a new frontier or near-frontier model every few days, and keep the evaluation habit that goes with that (see Picking the Right Model).

  2. Expect staged access for the most capable variants. The pattern this month was to ship first to a limited group, a cyber-partner program or trusted defenders, and later to everyone. When a launch post names a capability you want, check which tier gets it and when before building a client promise around it. The Astra rollout is covered in GPT-6; the earlier withholding decision is in OpenAI Withholds Astra.

  3. Assume the public model lags the lab’s best model. The reasoning comes from The Artificial Intelligence Show’s summary of Noam Brown’s interview with Dwarkesh Patel:

    • as agents take on longer tasks, “evaluating their behavior could take longer than the time between new model releases”;
    • so more safety time before release also widens the gap between what labs have and what the public can use.

    Paul Roetzer carries this to services: a lab could use unreleased models to “build businesses around AI that competitors cannot access”, naming law firms, consulting firms, medical practices and marketing agencies. For an agency, this is the scenario to watch rather than a slowdown of public releases.

  4. Keep a fallback that is outside any pacing agreement. Open-weight models from China, and any lab that declines to coordinate, are not bound by a US deal. An All-In host (machine-translated transcript) put it this way: “This stuff is out… You can download it on your computer.” A tested open-weight fallback for your core workflows protects you against a sudden access change. ^[inferred] See the Chinese open-weight decision framework.

  5. Watch liability, not the release calendar. The administration’s position is that labs keep full liability. If that holds, the practical effect is more gating, more refusals in sensitive domains and more verification requirements, rather than fewer releases. ^[inferred] The refusal trend is already visible in Opus 5.5 feedback:

    • one user benchmarking legal drafting against Fable 5.1 found Opus 5.5 “much worse due to increased safety rejections”;
    • a life-sciences team is still waiting on Anthropic’s verification program.

    Both are relayed by NLW; see Claude Opus 5.5.

Try It

  • Keep a two-line “model dependency” note for every client workflow: which model it needs, and what you would switch to if that model became gated, repriced or paused. Satya Nadella’s version of the test: “pull out a model and see whether I can retain the eval”, via Picking the Right Model.
  • When a launch says “rolling out to a limited set of organizations”, find out whether you are in the set before planning around it. Recheck the post a week later. Astra’s announcement promised availability to paid ChatGPT tiers “over the coming days” (Matt Wolfe), but the cyber-gated tiers have no public date.
  • Don’t sell a client on an unreleased capability you read about in a lab’s research post. The Navier–Stokes model was internal-only.

Open Questions

  • All quotes come from hosts reading on air. The Amodei and Pachocki essays, the Hubinger post and the resignation post were not fetched, and wording is from auto-captions, so it may be imprecise.
  • The resigning researcher’s name is captioned inconsistently (“Jacob Coxin”, “Coxon”). Berman calls him Anthropic’s “lead of alignment”, which conflicts with the Moonshots description (three years in pre-training at both labs) and with Hubinger holding the alignment science lead role.
  • The exact dates of the Pachocki and Amodei essays are inferred from episode references.
  • The scope and length of OpenAI’s pause. Berman describes it both as pausing “our latest RL run” and as “pausing AI development… a week and a half ago”. No source gives a start or end date.
  • Whether embedded evaluators add measurable lead time before release is unknown. Opus 5.5 shipped about ten days after the essay with external evaluation already done, so this first case suggests little delay.

Graph View

Backlinks