Model

Safety Was Anthropic's Whole Product. On Tuesday, Its Maker Walked Out.

A pretraining researcher quit the lab that sells caution, warning that both his old employer and his new one are “gambling with our lives.” The money did not blink.

Vincent JiangVincent Jiangfour-minute read
Share
Anthropic CEO Dario Amodei
Anthropic CEO Dario Amodei, from The Verge's Pentagon-terms coverage, 2026, a different outlet and a different story than the resignation sources. (Credit: The Verge)

Jacob Coxon resigned from Anthropic$350B — Anthropic, private, latest valuation $350B on Tuesday. His job was pretraining, the foundation work under Claude's models, first at OpenAI$1.18T — OpenAI, private, latest valuation $1.18T, then at the lab that markets itself as the careful one. Both employers, he now writes, are “racing straight to self-improving superintelligence and gambling with our lives.”1,2

What he actually said

In his exit note on X, he says the fear is not for show: “The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt.”1 His split of the two labs is blunt. OpenAI has not absorbed what it is building. Anthropic has absorbed it and builds anyway, on the theory that no rival will be careful, so it must win.2

He is not alone

He is not alone inside the building. A colleague, Evan Hubinger, posted publicly that risk from current models is low but compounding with “superintelligence arising from recursive self-improvement,” which is “happening faster than we thought.” Coxon calls for pacing agreements between US labs, or failing that, a temporary ban on capability improvements.2

The money did not blink

The money did not blink. OpenAI shipped its new Astra model the same morning.8 Legal AI startup Harvey$15.5B — Harvey, private, latest valuation $15.5B closed $550 million at $15.5 billion the same day.9 And Bill Gurley, echoed by Chamath Palihapitiya to 301,000 views, told the AI doomers to calm down and carry on, because things will be amazing.4,5

Why a board should care

Anthropic's commercial edge is trust, and trust is measurable. Claude has 245 million monthly users, third place, and converts 13% of them to paid subscriptions, the best rate in the industry.6,7

Chart of Anthropic's paid-conversion rate against rivals
Claude converts 13% of monthly users to paid, the best rate in the industry, and the exact asset a safety exodus puts at risk.

A safety brand that loses its safety researchers keeps the slogan and loses the substance.

The race has three scoreboards now: models shipped, capital raised, and researchers who leave. Watch the third one. It moves first.

Get Model by email, free

The AI labs, their models, and the business behind them. In our daily email, every weekday.

Reader comments

0 comments

    Model

    Hedra cut Krea 3's edit time from 22.9 to 3.3 seconds on AMD, and the app doesn't own the speed

    The eight-second edit Krea advertises is manufactured in the serving stack, and a partner's measurement shows the app does not own the speed. Inference efficiency accrues to the optimizer and the chip vendor, which caps a per-image app's margin and hands AMD a demand channel beyond hyperscalers.

    Model

    xAI ships AI teammates at $30-$200 a month, ten days after Musk said serving them loses money

    The published card sets a chatbot price against an always-on machine the CEO says loses money on every subscriber; SpaceX shareholders fund both sides of that trade in one stock.

    Model

    Google ships Gemini 4 Hydrogen at $1.50 a million tokens and gives it away in Search

    Gemini 4 Hydrogen scores 51 on the Artificial Analysis index against 50 for the flagship Argon, at $1.50 per million input tokens with no long-context surcharge, and it is free in Search. That strips both halves of a model lab's premium case the same week Anthropic prices an IPO on one.

    Model

    Scale AI just published the benchmark that prices its own Qatar contracts

    Scale AI's new Singapore legal benchmark measures exactly the gap that its Qatar contracts, signed three days earlier, pay Scale to close. Oracle's sovereign-AI story and private legal-AI valuations now have a number to argue with.

    Model

    Suno bought the top AI music spot with licenses; ACE Music v1 just took it, 1,160 to 1,144

    Suno bought the top of the AI music chart with licensed Warner, BMG and Believe catalogs; six weeks later a challenger holds it, and that challenger's model family is going free. The pricing test lands on Spotify's planned $4 to $5 remix tier and on what the majors can charge for AI training data.

    Model

    Midjourney says V9 could double its cost per image

    The self-funded lab is asking subscribers to accept twice the per-image cost for quality, the same week MiniMax open-sourced a 4K model whose license skips the US and the EU. The answer will price every credit-metered image AI, Adobe's included.

    Subscribe to TheInference for free