Anthropic Called This Model Too Dangerous to Release. Today It Put It on Sale.

This morning the company released Claude Fable 5, the first publicly available model in its Mythos class. Anthropic is blunt that Fable is the same underlying model as the Mythos it spent the spring warning the world about. The difference is what the safeguards will let it do. Ask Fable about cybersecurity, biology, chemistry, or model distillation and it refuses, routing the request to the older and weaker Opus 4.8 instead. Anthropic says it ran more than a thousand hours of bug-bounty and external red-team testing without anyone finding a universal jailbreak.

None of that is the question an enterprise should be asking. The real question is what you agree to the moment you turn it on.

Six Weeks From Forbidden to For Sale

Mythos launched as a preview in April, limited to a small group of partners under NDA on the grounds that wider access could lower the cost of cyberattacks and bioweapons. Last week, Anthropic widened access to hundreds of organizations across 15 countries. Today it opened a guardrailed version to anyone.

That sequence should give a buyer pause. A model billed as too dangerous to release in April, sold as safe enough for the public in June, is a remarkable thing to be able to walk a roadshow audience through. And there is a roadshow coming. Fable lands in the same quarter Anthropic filed confidentially to go public, days after OpenAI did the same.

Start With the Price, Because Price Is the Filter

Fable runs $10 per million input tokens and $50 per million output, roughly twice what Opus 4.8 costs. Opus was already the model making finance teams wince.

The timing is brutal. Companies are auditing AI spend right now and finding that reasoning-heavy models burn through tokens fast, splitting one request into a dozen hidden sub-tasks. Sam Altman has called AI budgets a "huge issue" for his customers, and we have written about why subsidized subscription pricing is a time bomb for the enterprises that built on it. A model at double the going rate does not ease that pressure. It compounds it.

There is also a near-term cliff buried in the rollout. Through June 22, Fable is included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost. On June 23, Anthropic pulls it from those plans, and continued use shifts to consumption credits. If your teams build a workflow around Fable over the next two weeks, the bill changes character on the fifteenth day.

For a narrow set of high-value autonomous work, the math can clear. Rakuten, an early customer, says the extra reasoning pays for itself. For ordinary knowledge work, you are paying a premium for headroom you may never use.

Read the Data Terms, Because They Just Changed

This is the line that should stop a CISO cold.

With Fable and Mythos 5, Anthropic now requires 30 days of retention on all traffic, overriding the zero-retention agreements some enterprises had already negotiated. The company says it will not train on that data and will use it only to catch novel attacks and reduce false refusals.

Take that at face value and it is still a problem. If you operate in healthcare, finance, or defense, or anywhere else that touches privileged or regulated data, a forced retention reversal is not a footnote. It is a compliance review, a renegotiated data-processing agreement, and possibly a legal no.

And because Anthropic sets the pace at the frontier, expect this to become the template. Access to the best models, bundled with retention rules sold as a safety feature. If you assumed your zero-retention posture was permanent, it just stopped being permanent.

The Guardrails Are a Capability You Are Paying to Lose

Anthropic says about 95 percent of sessions run start to finish on Fable, with only rare hand-offs to Opus. The company also concedes that some legitimate requests will get blocked, at least early on.

For most teams that is background noise, but for a security team or a pharma researcher, it is a deal breaker. You would be paying frontier prices for a model engineered to route your real work to a weaker one, with no clean way to separate a false refusal from a correct answer.

Our advice is to pilot the exact tasks you intend to run before you commit a budget to them. The benchmark you care about is your own.

The Benchmarks Are Strong. They Are Also Interested.

Anthropic points to outside results to make its case. An analytics platform reporting a score no prior model had reached. Coding and agent platforms reporting cleaner one-shot apps and better tool use.

Take those seriously, and take them as interested. They are customer testimonials packaged with the launch, not independent audits, and "did great in our evaluation" is the most flattering thing any vendor partner ever says. The genuinely external benchmarks measure raw capability, not the thing an enterprise is buying, which is reliable performance on its own workloads under its own constraints.

What Enterprise Leaders Should Be Doing Right Now

Do not start with whether Fable is the most powerful model available. Start with three quieter questions, and answer them before the June 23 pricing change forces your hand.

What does it cost per real workload, not per seat. Model the token burn on the tasks you would actually route to it, then compare against Opus 4.8 and against the model you run today.

What are the data terms at renewal, not just at signing. Get the 30-day retention requirement in front of legal and compliance now, especially if you hold a zero-retention agreement you assumed was durable.

How often will it refuse the work you bought it for. Run your real prompts, count the hand-offs to Opus, and decide whether the guardrail tax is one you can absorb.

The benchmarks make the slide. Those three answers make the decision.