Anthropic Launches Claude Fable 5, a New Tier Above Opus
Anthropic's Claude Fable 5 debuts a new Mythos class above Opus, pairing frontier capability with automatic safety routing for risky queries.
Anthropic did not bury the lede. On Tuesday the company introduced Claude Fable 5, and with it a label nobody outside the building had used before: Mythos class. Not Opus-plus, not Opus 5. A new rung altogether, sitting above the model family that has anchored Anthropic's lineup for the past two years.
The announcement, posted to Anthropic's newsroom, frames Fable 5 as the generally available half of a two-model release. Its sibling, Claude Mythos 5, ships the identical underlying architecture with the safety scaffolding stripped down for a narrow set of vetted users. Fable 5 is the version everyone else gets.
What makes a model Mythos class
Anthropic says the Mythos tier began quietly in April with a preview build, Claude Mythos Preview, offered by invitation only. Fable 5 and Mythos 5 are its public successors, and the company is pricing Fable 5 at ten dollars per million input tokens and fifty dollars per million output tokens, which Anthropic describes as less than half the cost of that earlier preview. It is still roughly double what Claude Opus 4.8 costs to run, a gap the company is betting customers will pay for the jump in capability.
According to coverage from Enterprise DNA, independent testers have already clocked Fable 5 around 80 percent on SWE-Bench Pro, a benchmark that scores how well a model can complete real, unassisted software engineering tickets. Anthropic's own examples lean the same direction: the company cites Stripe describing the model as having compressed months of engineering work into days, and highlights results in long-context reasoning, financial analysis, and even a demonstration of the model playing through Pokemon FireRed using nothing but screenshots as input.
The scientific claims are just as pointed. Anthropic says Fable 5 accelerated aspects of drug design workflows by roughly tenfold in internal partner testing and produced novel hypotheses in molecular biology, the kind of dual-use territory that immediately raises the question of who should be allowed to ask it what.
Automatic guardrails, not a refusal button
That question is where Fable 5's design gets interesting. Rather than a blanket refusal for sensitive topics, Anthropic built a classifier layer that watches incoming requests for three patterns: offensive cybersecurity tasks, dual-use biology or chemistry requests, and attempts to distill or extract the model's own capabilities. When one of those trips, the session gets quietly rerouted to the more conservative Claude Opus 4.8 instead of Fable 5 answering directly.
Anthropic says that reroute fires in fewer than five percent of sessions, and that external red-teaming logged more than a thousand hours of adversarial testing without producing what the company calls a universal jailbreak. Data touched by the safety classifiers is retained for thirty days for review purposes and then deleted automatically, a policy that applies specifically to Mythos-class traffic rather than Anthropic's broader retention rules.
Rolling out in phases, not all at once
Access did not flip on for everyone simultaneously. Fable 5 reached the Claude API and Amazon Bedrock immediately, while the rollout to Claude Pro, Max, Team, and Enterprise subscribers was staged between June 9 and June 23, with the model offered free during that introductory window according to reporting from Yahoo Finance.
Early hands-on impressions landed within hours. Developer and longtime Claude watcher Simon Willison, writing on his blog, described the model as "relentlessly proactive" in the techniques it reaches for when working toward a stated goal, a trait that reads as either impressive initiative or something to keep an eye on depending on the task you hand it.
Where Mythos 5 fits
Claude Mythos 5 is the half of the release most people will never touch directly. It runs on the same weights as Fable 5 but with the cybersecurity and biology classifiers selectively lifted for approved users, granted through a program Anthropic calls Project Glasswing. Cyber defenders who qualify get the cyber safeguards removed entirely; biomedical researchers admitted through a separate trusted-access track get the biology and chemistry restrictions lifted while the cyber protections stay in place. Anthropic frames this as narrowing the gap between what a model can theoretically do and what a legitimate specialist actually needs it to do, without handing that same headroom to anyone with an API key.
It is a more granular approach to model access than Anthropic has tried before, and it puts the company in the position of deciding, case by case, who counts as trusted enough for the unrestricted version. That framework was about to be tested sooner than anyone at Anthropic likely expected.
For now, the headline is simpler: there is a new ceiling above Opus, it has a name, and enterprise customers already running Fable 5 in production say the jump is not subtle. What happens when a model this capable meets the messier realities of government oversight and dual-use anxiety is the next chapter, and it started developing within days of launch.
Published in The Outspoken Digest
Read Next
More Technology →
