Claude Fable 5 and Mythos 5 — Anthropic Ships Its Most Capable Public Model With a 'Safeguard That Routes Back' Design (2026/06/09)

Anthropic Model Release 2026/06/09

Anthropic Official Announcement · 2026-06-09 "For queries on certain topics, the next most capable model, Claude Opus 4.8, responds instead — and it is with such safeguards in place that we have released this model."

On June 9, 2026, Anthropic released the same underlying model under two names. The generally available Claude Fable 5 (model ID claude-fable-5) and the limited-availability Claude Mythos 5 (claude-mythos-5). The two are identical inside; the only difference is "the presence or absence of safeguards." Fable 5 claims to be the company's most capable public model, while it automatically routes requests flagged in the cybersecurity / biology and chemistry / model distillation domains back to Claude Opus 4.8. The safeguard-stripped Mythos 5 reaches only a limited set of cyber defenders via Project Glasswing.

What characterizes this announcement is that it splits capability, safety, and access into separate levers. Even for the same model, if the safeguard classifier raises a flag, the request is routed back to a lower model, while the version with safeguards removed is handed only to vetted parties. "Ship the most capable public model" and "narrow down dangerous uses" are reconciled not in the model itself but in the surrounding routing and distribution.

Fable 5's Capabilities

Anthropic positions Fable 5 as "state-of-the-art on nearly every evaluation benchmark," claiming high performance in software engineering, knowledge work, vision, and scientific research. In third-party evaluations, it is reported as "the highest score among frontier models even at medium effort" on Cognition's FrontierCode, "the highest of any model" on Hebbia's finance benchmark, and "almost fully solving" IMC's trading-analysis evaluation (all based on Anthropic's announcement).

As a concrete example, the Stripe case is cited. On a 50-million-line Ruby codebase, a whole-codebase migration that would normally take an all-hands team more than two months was completed in one day, said to have "compressed a months-long process into days." It is also said to be more token-efficient than past Claude. The length of autonomous operation is emphasized as well: "it can work autonomously longer than any Claude before," and in the game Slay the Spire, file-based persistent memory is said to have boosted performance to "3x Opus 4.8." In vision, it is reported to have extracted accurate numbers from scientific charts, reconstructed the source of web apps from screenshots, and completed Pokémon FireRed with a minimal vision-only harness.

Safeguards — "Route Back to the Next Most Capable Model"

The design's core in Fable 5 is automatic fallback (classifier-based routing) . In three domains — cybersecurity, biology and chemistry, and model distillation — requests flagged by the safety classifier are answered not by Fable 5 but by the next most capable model, Claude Opus 4.8. Anthropic states this triggers "on average in under 5% of sessions," explaining that in over 95% of sessions no routing-back occurs.

The reasoning is made explicit as well. On cyber, Mythos-class models are "adept at finding and exploiting software vulnerabilities" and "could make cyberattacks substantially cheaper and easier." In biology and chemistry, the ability to carry out real-world scientific tasks has risen, and while citing an example of surpassing a specialized model in adeno-associated virus (AAV) design, "the same capability in the wrong hands could enable the design of dangerous viruses." On distillation, attempts "to extract Claude's capabilities at scale to train competing models in authoritarian states" create the risk that "near-frontier capability spreads without appropriate safeguards." Anthropic acknowledges that "the safeguards are intentionally tuned to the cautious side and are still stricter than ideal," and says it will reduce false positives after launch.

Developer-side behavior changes too. A refusal returns in the API not as an error but as HTTP 200 + stop_reason: "refusal", indicating which classifier refused. With the fallbacks parameter (beta) or SDK middleware, you can automatically retry to a different model; requests refused before output generation are not charged, and on retry the cost of switching prompt-cache is refunded. Note that Fable 5 / Mythos 5 are designated as Covered Models, subject to 30-day data retention, and cannot be used under zero data retention.

Mythos 5 and Project Glasswing

Claude Mythos 5 is "the same underlying model as Fable 5, but with safeguards removed in some domains." It is provided via Project Glasswing, a collaboration with the US government, and is the successor to Claude Mythos Preview. Anthropic claims Mythos 5 "has stronger cybersecurity capabilities than any model in the world" (the company's assertion). Initial access is limited to a small group of cyber defenders and infrastructure providers, currently in the form of Glasswing partners upgrading from Preview. Planned ahead are a broader trusted access program through which cyber organizations can systematically apply, and a separate program that opens a version with only the biology and chemistry safeguards removed (cyber safeguards retained) to a subset of life-science researchers.

The scientific results shown with Mythos 5 are also included in the announcement. In drug discovery, in-house protein specialists "accelerated part of the design process by roughly 10x," obtaining strong candidates for 9 of 14 protein targets. On molecular-biology hypotheses, researchers "preferred Mythos's hypotheses about 80% of the time versus the Opus class," and one hypothesis was reportedly corroborated by a lab independently working on the same problem. In genomics, with over a week of largely autonomous work, it is reported to have built a custom ML model that "outperforms a recent model published in the journal Science at 1/100th the scale" (all Anthropic's assertions).

The origin of the naming is appended as well. "Fable derives from the Latin fabula (= that which is told), close to the Greek mythos. What separates the two models is the safeguards." A design philosophy of giving the same contents two names and distinguishing them by the presence or absence of safeguards alone is folded into the names themselves.

Pricing and Availability

Pricing is $10 per million input tokens and $50 for output. Exactly double Claude Opus 4.8 ($5 input / $25 output), and said to be less than half the price of the old Mythos Preview. Context is 1 million tokens by default, with a maximum output of 128k tokens. adaptive thinking always on is the only thinking mode; thinking cannot be turned off, and its depth is adjusted via effort. The raw thinking process is not returned; only a summarized version can be obtained. At launch it supports effort, task budgets (beta), the memory tool, context editing (beta), compaction, and vision.

Availability begins June 9, 2026. Fable 5 is generally available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry. It is usable from day one on usage-based Enterprise as well. For subscriptions (Pro / Max / Team / seat-based Enterprise), June 9–22 is included at no additional cost, and from June 23 onward usage credits are required, with the stated aim of returning it to standard availability as capacity permits. Mythos 5 is limited to Glasswing partners and is not generally available.

Editorial Observations

The most readable aspect of this release is its distribution design — "splitting capability, safety, and access into separate levers" — rather than the model's capability itself. The frontier safety problem, which previously tended to look like a binary of "ship a strong model or don't," Anthropic decomposed into a two-stage structure: (1) embed classifier routing in the public version to route only the dangerous domains back to a lower model, and (2) hand the same model with safeguards removed only to vetted cyber defenders via Glasswing. The naming that distinguishes the same contents under the two names Fable and Mythos is a statement of that philosophy.

What matters from the MEMEX perspective is that this shows, in primary sources, a shift from "draw a red line and stop" safeguards to "ship the capability, but narrow uses via routing and distribution." Anthropic's own stance of speaking — "the safeguards are stricter than ideal," "we will reduce false positives" — on the premise that false positives exist is also an open acceptance that the safeguards come with operational friction (= under 5% of sessions drop to a lower model). The "business structure that can say No" seen in Dario's clash with the Pentagon and this "ship the capability but narrow via routing" design can be read as two faces of the same company's safety operations.

Points of Focus

Classifier routing as a new safety primitive

The mechanism of "route back to the next most capable model when a flag is raised in a dangerous domain" is a third path that neither neuters the model at the training stage nor refuses requests uniformly. In the API, a refusal returns as HTTP 200 + stop_reason: refusal, with disclosure extending to which classifier refused. By placing safeguards not "inside the model" but in "a routing layer around the model," the structure makes it possible to independently adjust public capability and the control of dangerous uses. The figure of triggering in under 5% of sessions also suggests an intent to quantify and display the scale of the friction.

One model, two names — only the safeguards differ

The fact that Fable 5 and Mythos 5 are technically the same contents, differing only in the presence or absence of safeguards, expresses the position that "a model's danger" is not a fixed attribute but a relative one that changes with distribution conditions. Removed for cyber defenders; routed back for the public. For biology and chemistry, removed only for a subset of researchers through yet another trusted access program. While keeping capability constant, a distribution that finely carves out "to whom, in which domain, and how far" is becoming the new way to ship frontier models.

Related Resources

Glossary

Claude Fable 5: A public model released by Anthropic on 2026-06-09, claimed to be the company's most capable (model ID claude-fable-5). 1M context, max output 128k tokens, pricing of $10 input / $50 output per million tokens. It embeds a safeguard that routes requests flagged in cybersecurity / biology and chemistry / model distillation back to Claude Opus 4.8.
Claude Mythos 5: A limited-availability version that is the same underlying model as Fable 5 but with safeguards removed in some domains (claude-mythos-5). Provided only to a small group of cyber defenders and infrastructure providers via Project Glasswing, a collaboration with the US government. The successor to Claude Mythos Preview. Anthropic claims "stronger cybersecurity capabilities than any model in the world."
Automatic fallback (classifier routing): Fable 5's safety mechanism. Requests that a safety classifier flags in a dangerous domain (cyber / biology and chemistry / distillation) are automatically routed back to the next most capable model, Claude Opus 4.8. Triggering occurs in under 5% of sessions. In the API, a refusal returns as HTTP 200 + stop_reason: refusal, and if no output was produced before the refusal, no charge is incurred.
Project Glasswing: A collaboration framework Anthropic is advancing with the US government. A channel that provides safeguard-stripped Mythos-class models only to vetted cyber defenders and infrastructure providers. Mythos 5 is distributed through this framework as the successor to Claude Mythos Preview.
adaptive thinking always on: The only thinking mode in Fable 5 / Mythos 5. Thinking cannot be disabled, and its depth is controlled by the effort parameter (low to max, default high). The raw chain-of-thought is not returned; only a summarized version can be obtained. The Messages API behavior of Opus / Sonnet / Haiku is unchanged; this change is specific to Fable 5 / Mythos 5.