Anthropic's Safety Story Is Its Business Strategy. The Numbers Prove It.

Anthropic has spent three years building the most expensive narrative in AI: that it is the responsible lab. This week, Ben Thompson’s Stratechery analysis argued the same thing Thompson has been arguing for months — but with three concrete pieces of evidence that turn the framing from a moral stance into a recognisable business strategy. The economic imperative, the data imperative, and the power imperative all point the same way. Safety is not a side effect. It is the moat.

🔍 THE BOTTOM LINE

Anthropic’s safety positioning does three jobs at once: it justifies the prices Anthropic charges above cost, it legitimises the data-collection changes that make the next model better, and it positions the company as the gatekeeper that any future AI regulation has to go through. None of this requires Anthropic to be insincere. Thompson is careful to say safety is a genuine belief system for the people who run the company. The argument is that it is also, simultaneously, a competitive strategy — and that understanding both is the only way to read the next twelve months of Anthropic’s behaviour.

What the Fable/Mythos Release Triggered

Two months ago, Anthropic announced Mythos was too dangerous to release publicly because of its advanced cybersecurity capabilities. The version Anthropic did release — Fable 5 — was, in Thompson’s words, “extremely impressive” enough to make GPT 5.5 and Opus 4.8 “feel small and dumb.” That matters because only GPT-4 and Grok 4 had previously created that sensation, and both were generational jumps in base model size. Fable appears to be the first of a new generation.

Then a jailbreak was discovered. Reports — including from Axios — indicate the vulnerability was reported by Amazon, which is simultaneously an Anthropic investor and one of the largest inference providers for the model. The U.S. government issued an export-control directive suspending Fable 5 and Mythos 5 access for any foreign national, anywhere in the world. We covered the directive on June 14 and Andy Jassy’s role in escalating it on June 15. Anthropic called the move a “misunderstanding.” White House officials called Anthropic’s response “insouciance by the company’s leadership to legitimate national security concerns.” Senior staff are now in Washington.

The Three Imperatives

Thompson identifies three pressures that all push Anthropic toward the same strategy. They are independent. They all reach the same conclusion.

1. The Economic Imperative

Compute value has flown to Nvidia, TSMC, SK Hynix, Samsung, and Micron. The labs themselves are losing money. Thompson writes that Anthropic and OpenAI have “collectively lost tens of billions of dollars” on models that are then distilled by open-source competitors — primarily in China — into “good enough” alternatives that undercut on price. The labs have a problem: the differentiation is fleeting.

The strategic response is to own the user touchpoint. The most valuable place in the value chain has always been the user touchpoint, and that is now where the AI companies have to be — on a collision course with software companies that have historically owned the touchpoint. Microsoft’s Satya Nadella put a counter-vision on X earlier this year: companies should build “token capital” they own, and be able to swap the underlying model without losing the institutional knowledge. “You can offload a task, or even a job, but you can never offload your learning,” Nadella wrote, and warned: “The last thing any of us want is a world where every company across every sector is ceding value to a few models that eat everything they see. If all the value is accrued by only a few models, the political economy will simply not tolerate it.”

Thompson’s read is that the globalization analogy is not a warning but a prophecy: the hollowing out of industrial economies happened because globalisation did exactly what Nadella fears. “The economic imperative for the model makers is to accomplish exactly this,” Thompson writes. Anthropic’s safety framing is what allows it to position itself as the responsible owner of the touchpoint — the one governments, enterprises, and users can trust to handle the value concentration.

2. The Data Imperative

Model improvements now come from reinforcement learning on real-world use. Synthetic data has limits. The real lever is what users actually do with the model, which means the more interactions an AI company has, the better its next model gets. This is the unspoken reason for the heavily subsidised enterprise plans: SemiAnalysis estimates that a $200-a-month plan corresponds to roughly $8,000 of Claude tokens and $14,000 of Codex tokens at list price. Anthropic is selling to enterprises at a fraction of the cost because the real return is the data those enterprises generate.

The Fable 5 release came with a quiet but consequential change: Anthropic now retains all usage data for 30 days, including for enterprise plans that previously had zero retention. Anthropic says it will not train on the data. The retention window is the new bit. For a model whose improvements are increasingly driven by real-world RL, 30 days of corporate usage — even if not used directly for training — is a meaningful operational asset. The framing is safety. The mechanism is data capture.

3. The Power Imperative

By being the lab that says “this model is too dangerous to release,” Anthropic positions itself as the entity that gets to decide what is and is not safe. When the U.S. government disagreed, the conflict was over who controls that gate. The directive suspending Fable 5 and Mythos 5 is in part a power move by the state to assert that it — not Anthropic — decides which AI capabilities can leave U.S. jurisdiction. The fact that senior Anthropic staff are now in Washington is the inverse of the same negotiation: the company is trying to retain enough standing with the state to keep being the gatekeeper, even as the state pushes back on which decisions the gatekeeper can make unilaterally.

The combination — economic moat, data moat, regulatory access — is what Thompson is calling the safety superpower. None of it works without the framing. The framing is what gets Anthropic the enterprise contracts at premium prices, the regulatory capture that shuts out smaller competitors, and the trust of users who would otherwise use a distilled open-source model. The same sentences appear in the same press releases for all three reasons.

The Nadella Counter-Vision, and What It Means for NZ

Microsoft’s Nadella is, in Thompson’s framing, offering a different answer to the same problem. Build the “token capital” inside the company. Train the model on your own data. Own the institutional knowledge. The model becomes a commodity input that can be swapped — the way electricity can be sourced from any generator on the grid. Value stays with the enterprise, not with the model provider.

For New Zealand companies, this distinction is sharper than for U.S. firms. NZ has no domestic frontier AI lab. The choices for a Kiwi enterprise are essentially: pay Anthropic (or OpenAI, or Google) for a model that will improve itself on your data, or build the data layer yourself and accept that you are running a model someone else trained. The sovereign-AI question lands harder here because the answer is not “we’ll build our own ChatGPT” — it is “we’ll build our own institutional data layer, and we’ll choose a model that we trust not to extract the value.”

The first approach is faster. The second is the only one that survives the next round of price increases, retention policy changes, or export-control directives.

❓ FAQ

Is Anthropic’s safety talk just marketing, then? Not according to Thompson’s argument, and not according to the people who run the company. The point is that safety can be both a genuine belief and a competitive strategy at the same time. The two are not mutually exclusive. Treating the safety framing as purely cynical misses the operational reality; treating it as purely altruistic misses the strategic logic. The same actions are both.

What did the 30-day data retention change actually do? Before Fable 5, enterprise plans on Anthropic had zero data retention by default. After the change, all usage data is retained for 30 days regardless of plan tier, though Anthropic says it will not train on the data. The retention window is the new bit, and it gives Anthropic 30 days of high-value enterprise usage to draw on for safety monitoring, incident response, and (potentially) future product development.

How does this connect to the export-control fight? The export-control directive against Fable 5 and Mythos 5 is the state asserting control over the gate. Anthropic wants to be the gatekeeper because that is where the value concentrates. The U.S. government wants to be the gatekeeper because AI capability is now a national-security asset. The D.C. negotiation is over who gets to make which decisions about which models reach which customers.

What is “token capital”? A term Satya Nadella has used to describe the proprietary knowledge a company builds into its AI workflows over time. The point is that the model is interchangeable; the company’s data, context, and institutional patterns are not. If a company builds its token capital well, it can switch model providers without losing the value of what it has built.

Is any of this a problem for an individual Kiwi using Claude? At the consumer level, no — the Fable 5 export-control directive means Fable 5 itself is no longer available to foreign nationals anywhere, so individual users are not currently affected by the latest model release. The strategic question is for the enterprises and government agencies making multi-year procurement decisions about which AI provider to standardise on.

🔍 THE BOTTOM LINE

Anthropic’s safety story is not a lie and it is not a side effect. It is the product. The export-control fight, the 30-day retention change, the $200 plans that cost Anthropic $22,000 in tokens — all of these are the visible artefacts of a coherent strategy. The strategy works as long as the safety framing is trusted by enough enterprises, governments, and users to keep the value concentration flowing toward the gatekeeper. The question for 2026 and 2027 is whether the trust holds — and what the world looks like if it doesn’t.

Singularity.Kiwi

Anthropic's Safety Story Is Its Business Strategy. The Numbers Prove It.

🔍 THE BOTTOM LINE

What the Fable/Mythos Release Triggered