Anthropic makes safety a power claim

Ben Thompson’s Stratechery essay frames Anthropic’s Fable and Mythos dispute as a preview of the next fight in AI: who controls frontier models, customer data, and the software layer that sits between companies and users.

Anthropic, according to Thompson, faces the same business pressure as OpenAI and other frontier labs. Model companies spend billions on training runs, then watch open-source and lower-cost models copy much of their capability. That pushes labs toward the user interface, where products can create lock-in and collect usage data.

The Fable launch shows that pressure. Anthropic released a guarded version of Mythos after saying Mythos carried cybersecurity risk. The U.S. government then ordered Anthropic to suspend foreign-national access to Fable 5 and Mythos 5, according to the company’s account, after officials reviewed a jailbreak technique. Anthropic argued the technique found minor, known vulnerabilities that other models could also identify.

Thompson does less with the disputed facts than with the incentive structure. If Fable does not warrant a government fight, he argues, a later model will. Frontier labs now use models to help build stronger successors. That loop gives national security officials a reason to intervene, and it gives Anthropic a reason to insist that its safety process should govern access.

The economic fight sits behind the safety fight. Microsoft CEO Satya Nadella has warned that companies should build their own “token capital” instead of handing value to a few general models. His position protects enterprise software companies that want models to serve as replaceable inputs. Anthropic and OpenAI have the opposite incentive: they gain more value when users work inside their products instead of inside legacy software.

Data makes that conflict sharper. Thompson points to Anthropic’s 30-day retention policy for Fable usage, including enterprise usage that once carried zero-retention promises. Anthropic said it would use the data for safety monitoring, not training. Thompson argues the same data would help Anthropic improve its models and tighten the product loop: more user workflows produce more traces, and better traces help labs improve the model.

Anthropic’s proposed restrictions on frontier large language model development added a second concern. The company said Fable would reduce its usefulness for certain requests tied to competing model development, then said it would instead hand those requests to Opus 4.8 with disclosure. Thompson reads the first plan as a sign that Anthropic wants to decide which competitors deserve model help.

That claim fits Anthropic’s founding story. The company broke from OpenAI around safety concerns and built its identity around alignment. Thompson’s strongest point comes from that fit between mission and business. Anthropic can tell researchers they are building powerful AI under a safety mandate, tell customers it needs data access for abuse detection, and tell governments it deserves room to manage dangerous capability.

The piece treats that alignment as Anthropic’s advantage and risk. A company with a clear mission can move faster than a rival caught between research culture and consumer scale. The same conviction can also lead executives to treat their own judgment as a substitute for customer choice, regulatory oversight, or market limits.

For buyers, the lesson is practical. A frontier model contract now covers more than price, latency, and benchmark scores. Companies must read data retention terms, model substitution rules, usage restrictions, and audit rights with the same care they bring to cloud infrastructure agreements. The vendor that powers an internal workflow may also learn from that workflow, limit certain uses, and steer customers toward its own product surface.

For the AI market, Thompson’s essay identifies the pressure line between model makers and software companies. Labs need user data and product control to escape commodity pricing. Software companies need model interchangeability to preserve customer ownership. Safety gives Anthropic a language for its choices, but customers and governments will judge the choices by who gains power from them.