Anthropic shipped both ends of the curve on one day

Ronni Holmvig Strøm · 2026-07-02

On June 30, Anthropic did two things that look unrelated and are not. It released Claude Sonnet 5, a cheap and aggressively agentic model that it will make the default for every free and paid

On June 30, Anthropic did two things that look unrelated and are not. It released Claude Sonnet 5, a cheap and aggressively agentic model that it will make the default for every free and paid user. And it won back Fable 5, its most capable public model, which the US Commerce Department had forced offline on June 12 over cybersecurity fears. One is a commodity you can rent for two dollars a million tokens. The other is closer to a controlled export. They are the same capability curve, photographed at both ends on the same day.

Safety Is Now a Product Feature

Sonnet 5 can make plans, drive browsers and terminals, and run autonomously at a level that, a few months ago, needed larger and more expensive models. Anthropic is making it the default for free and Pro accounts and pricing it at $2 per million input tokens through August, well under Opus 4.8, OpenAI's GPT-5.5, and Gemini 3.1 Pro. Autonomous agentic work, the thing every lab was still charging a premium for at the start of the year, is now the entry tier.

What is striking is how much of the pitch is about restraint. Anthropic says Sonnet 5 refuses malicious requests more cleanly and resists prompt-injection hijacks better than its predecessor. It even checks its own output without being asked. And in the blog post's own words, it has "much lower ability to perform dangerous cybersecurity tasks than our current Opus models."

Read that again. A frontier lab is now advertising what its model deliberately cannot do, and selling it as a reason to buy. Fabian Hedin, co-founder of Lovable, put the logic plainly: "A model that knows when to say no is just as important as one that knows how to build." Capability at the low end has become table stakes. Safety is the thing worth naming.

The Frontier Now Moves Like a Controlled Substance

Now the other end. Fable 5 and its more powerful sibling Mythos 5 debuted on June 9 as the most capable systems Anthropic has ever released to the public. Fable 5 shipped with guardrails blocking a range of cyber and biology tasks. Senior administration officials worried those guardrails could be circumvented, and though experts disagreed over how serious the risk really was, Commerce Secretary Howard Lutnick enacted export controls barring foreign nationals from the model. To comply, Anthropic pulled both models offline on June 12.

Then it sent a team of its top scientists to Washington. Nineteen days later, the controls came off. Anthropic had built a new method of blocking the specific cyber workaround that worried officials, and Commerce's own Center for AI Standards and Innovation tested and backed it. Fable 5 began rolling back out on July 1. Mythos 5 returned to about 100 trusted organizations working on cybersecurity and infrastructure.

The episode carried a real cost. Some foreign organizations that had access before June 12 are still locked out, Asian startups shipped Mythos-like models while the ban dragged, and Sam Altman called the emerging pattern of phased, government-approved launches "bad news" for the field. These are the frictions of a country working out, in real time, how to govern a capability that has outgrown a hands-off posture.

Capability and Oversight Are Converging

The dangerous-cyber capability that got Fable 5 export-controlled in June is, one tier down and a few months on, exactly the capability Sonnet 5 ships cheaply and boasts about lacking by design. The frontier and the commodity are the same engineering, arriving at different speeds, and the safety work at the top eventually becomes the default at the bottom.

The nineteen-day shutdown did not end with a model buried or a lab broken. It ended with better guardrails, vetted by government experts, and the models back online.