Claude Fable 5: Anthropic unlocks Mythos-class AI for everyone

On June 9, 2026, Anthropic announced the release of Claude Fable 5 — the first Mythos-class model available to the general public. It is the same core model that was previously restricted to a small group of organizations under the confidential Project Glasswing, now wrapped in a new classifier layer that blocks the highest-risk use cases. In parallel, the company is releasing Claude Mythos 5 — a version without cybersecurity restrictions — but only for verified government partners and researchers.

Key takeaways

Fable 5 is the first Mythos-class model in general availability — the same core as Mythos Preview, with an added classifier layer
Pricing: $10/1M input tokens and $50/1M output tokens — more than double the rate for Claude Opus 4.8
SWE-bench Pro: 80.3% (Fable/Mythos 5) vs 58.6% GPT-5.5 and 55.2% Claude Opus 4.8
Classifiers block three domains: cybersecurity, biology and chemistry, model distillation
Available on Pro/Max/Team plans at no extra cost until June 22; usage credits required from June 23

How the model works and how it differs from Mythos 5

Fable 5 and Mythos 5 share the same base model — the difference is purely access control. When Fable 5's classifiers detect a query related to cybersecurity, biology or chemistry, or model distillation, the system automatically routes the response to Claude Opus 4.8. Users are notified when this happens. Anthropic says more than 95% of sessions run with no fallback at all.

Mythos 5, available only to Glasswing partners and soon to select biology researchers, has its cybersecurity restrictions lifted (and partially its biology restrictions). On ExploitBench it scores 78.0% vs 34.0% for GPT-5.5, which Anthropic calls "the strongest cybersecurity capabilities of any model in the world."

The naming is deliberate: fable comes from Latin fabula — that which is told. Mythos is the deeper layer: the same model, without filters. Both carry the number "5" referring to the Mythos Preview generation, not a sequential numbering within the Claude family.

Benchmarks: code, finance, vision

In software engineering, Fable 5 dominates especially on long-horizon tasks. On FrontierCode Diamond — a benchmark testing production-quality agentic coding — the model scores 29.3%, while Opus 4.8 reaches 13.4% and GPT-5.5 just 5.7%. Stripe tested the model on a 50-million-line Ruby codebase and reported that a full codebase migration that would normally take a team over two months was completed by Fable 5 in a single day.

In analytical knowledge work, the model was the first to break 90% on Hex's core analytics benchmark — a 10-point jump over Opus 4.8. On GDPpdf (reasoning over PDF documents with charts and tables), Fable 5 scores 29.8% without tools, while GPT-5.5 stops at 24.9%.

Vision is another clear strength. The model can reconstruct a web application's source code from screenshots alone and completed Pokémon FireRed using only raw game screenshots with no additional tooling — something previous Claude versions could not do even with a full helper harness.

Safeguards — three risk categories

Anthropic kept Mythos Preview behind closed doors for months before this release, citing the model's ability to find and exploit software vulnerabilities. Fable 5 addresses this not by fully blocking these capabilities but through routing: sensitive queries go to Opus 4.8, everything else goes to Fable 5.

Three domains covered by classifiers:

Cybersecurity. Mythos-class models can autonomously carry out agentic hacking tasks: from reconnaissance through lateral movement to vulnerability exploitation. External evaluation found zero successful jailbreak attacks using 30 known public techniques. However, UK AISI made progress toward a "universal jailbreak" in a brief testing window — Anthropic acknowledges that fully preventing this is likely impossible; the goal is to increase cost and detectability.

Biology and chemistry. Mythos-class models outperform dedicated protein language models on tasks involving the design of adeno-associated virus (AAV) shells. The same capabilities that accelerate gene therapies can be misused for harmful biological work.

Model distillation. Anthropic has previously identified attempts to extract Claude's capabilities at scale to train competing systems — including in authoritarian countries. Classifiers are designed to block this for Fable 5.

Data retention policy

As a new requirement, Anthropic is introducing 30-day data retention for all traffic on Mythos-class and higher models — for both first- and third-party surfaces. The company states the data will not be used to train new models or for non-safety purposes. For enterprises in regulated industries (finance, healthcare, legal), this is a meaningful policy change to evaluate.

Why this matters

The Fable 5 launch is the first public demonstration of a new frontier-model distribution mechanism: one core, different access layers depending on trust and domain risk. Rather than a simple "public vs. private" split, Anthropic is building a spectrum: Fable 5 for everyone, Mythos 5 for verified government partners, and eventually for biologists and cybersecurity firms.

For the industry, this is a significant observation. Models are becoming capable enough that no single policy approach is sufficient: full openness creates risk, full restriction hampers applications with enormous scientific and economic value. Routing to a less capable model in sensitive contexts is a compromise that lets Anthropic move faster than if it had to achieve perfect safeguards before any public release.

Competition will respond quickly: OpenAI and Google have their own models at similar capability levels. The question is no longer "who has the strongest model" but "who has the best mechanism for controlled distribution."

What's next

The free-access period for Pro/Max/Team plans ends June 22, 2026 — usage credits required from June 23; Anthropic commits to restoring Fable 5 as a standard plan feature "as quickly as sufficient capacity allows"
A trusted access program for biologists and biomedical companies is expected to launch within weeks — providing access to Fable 5 with biology restrictions lifted (while cyber restrictions remain in place)
Anthropic announced upcoming publication of genomics research conducted by Mythos 5 — the model autonomously analyzed single-cell data for millions of cells across 138 animal species and trained its own ML model on the results