Claude Fable 5 and Claude Mythos 5

Overview

Anthropic launched two flagship AI models on June 9, 2026: Claude Fable 5, made safe for general public use, and Claude Mythos 5, a restricted version released first to cyberdefenders. Both share the same underlying model, the company's most capable to date, and both are priced at $10 per million input tokens and $50 per million output tokens, less than half the price of the earlier Claude Mythos Preview. Fable 5 ships with safeguards that route sensitive queries to the older Claude Opus 4.8 model, while Mythos 5 has those safeguards lifted in some areas for trusted partners.

Key Takeaways

Claude Fable 5 is the public model and Claude Mythos 5 is the same model with safeguards removed in select areas, released first to cybersecurity defenders and infrastructure providers.
Both models cost $10 per million input tokens and $50 per million output tokens, less than half the price of the previous Claude Mythos Preview.
Fable 5 ships with safeguards that redirect risky queries on cybersecurity, biology, chemistry, and model-copying topics to the older Claude Opus 4.8 model.
The safeguards trigger in fewer than 5% of sessions on average, tuned conservatively so they sometimes catch harmless requests.
Mythos 5 launches through Project Glasswing, a US government collaboration, and Anthropic says it has the strongest cybersecurity capabilities of any model in the world.
Named partners including Stripe, Hebbia, Cognition, IMC, GitHub, and Cursor reported state-of-the-art results during early testing.

Stats & Key Facts

#Pricing set at $10 per million input tokens and $50 per million output tokens, less than half the cost of Claude Mythos Preview.
#Safeguards trigger in fewer than 5% of sessions on average.
#Stripe ran a codebase-wide migration on a 50-million-line Ruby codebase in one day, a task estimated at over two months of manual team work.
#Fable 5 broke 90% on Anthropic's core analytics benchmark, a 10-point gain over Opus.
#Scientists preferred Mythos 5's molecular biology hypotheses about 80% of the time in blinded comparisons against Opus-class models.
#Over 1,000 hours of external red-teaming produced no universal jailbreaks of the safeguards.

Two Models, One Engine: How Fable 5 and Mythos 5 Differ

The two releases run on the same underlying system but serve different audiences.

Claude Fable 5 is the version Anthropic made safe for general use. The company describes it as exceeding the capabilities of any model it has previously released to the public, with state-of-the-art results across software engineering, knowledge work, vision, and scientific research.

Claude Mythos 5 is the identical underlying model with safeguards lifted in some areas. Because removing those guardrails raises the stakes, access stays limited to vetted partners rather than the open public. Anthropic frames the dual launch as a way to ship advanced capability quickly while keeping the riskiest uses behind trust controls.

The naming swaps the usual pattern: Mythos is the more capable, less restricted track, while Fable is the safety-tuned public track built on the same foundation.

Conservative Safeguards Reroute Risky Queries to Opus 4.8

Anthropic built Fable 5 with a fallback system rather than relying on the model alone to refuse harmful requests.

›Queries flagged in three areas, cybersecurity, biology and chemistry, and attempts to copy the model, get answered by the older Claude Opus 4.8 instead.
›The safeguards are tuned conservatively, so they sometimes catch harmless requests.
›On average they trigger in fewer than 5% of sessions, meaning most users never hit them.
›Anthropic says it is working to reduce false positives as more capable models arrive in the coming months.

Project Glasswing and the Cyberdefense Mission for Mythos 5

Mythos 5 reaches a narrow group first through a government partnership.

Mythos 5 deploys initially through Project Glasswing, a collaboration with the US government, as an upgrade to the earlier Claude Mythos Preview. Anthropic states it has the strongest cybersecurity capabilities of any model in the world.

The company says the models have already helped cyber defenders secure critically important software during the preview phase. It plans to widen access to Mythos 5 later through a broader trusted access program rather than a public release.

The logic is straightforward: the same skills that help defenders find and fix security holes could help attackers, so Anthropic limits who gets the unguarded version.

Software Engineering Results from Stripe, Cognition, GitHub, and Cursor

Early enterprise testers reported large productivity gains on real codebases.

›Stripe ran a codebase-wide migration on a 50-million-line Ruby codebase in a single day, work it estimated would have taken a team over two months by hand.
›On Cognition's FrontierCode evaluation, which tests production-quality coding, Fable 5 scored highest among frontier models, even at medium effort.
›GitHub reported a level of autonomy and reliability in testing that exceeded previous benchmarks.
›Cursor called it a state-of-the-art model on its CursorBench evaluation.
›Anthropic notes Fable 5 is more token-efficient than past Claude models, meaning it does more work per unit of compute.

Knowledge Work, Vision, and Memory Improvements

The gains extend beyond coding into analysis and perception tasks.

On Hebbia's Finance Benchmark for senior-level reasoning, Fable 5 posted the highest score of any model, with gains in document-based reasoning and chart and table interpretation. Trading firm IMC said the model aced its trading-analysis evaluations nearly across the board. Fable 5 also broke 90% on Anthropic's core analytics benchmark for complex, long-running tasks, a 10-point jump over Opus.

For vision, Anthropic calls Fable 5 its new state-of-the-art model, noting it beat a benchmark called FireRed using a minimal, vision-only setup where earlier models needed elaborate helper systems.

Memory also improved. When the model played the strategy game Slay the Spire using file-based notes, persistent memory helped its performance three times more than it helped Opus 4.8, a sign of stronger long-running focus.

Life Sciences: Drug Design and Genomics Acceleration

Anthropic highlights early scientific use as a reason these models matter.

›Internal protein experts said the model accelerated parts of the drug design process by around ten times.
›Of 14 protein targets tested, nine produced strong candidates for drug design.
›Scientists preferred Mythos 5's molecular biology hypotheses about 80% of the time in blinded comparisons against Opus-class models, and one hypothesis was later corroborated by an independent lab.
›In a week of autonomous genomics research, Mythos 5 assembled single-cell data across 138 species and built a model 100 times smaller that outperformed published Science journal work.

Safety Testing and What It Means for Business Readers

Anthropic published red-teaming results to support its safety claims.

External red-teaming ran for over 1,000 hours and found no universal jailbreaks of the safeguards. One outside partner judged Fable 5's safeguards the most robust of any model tested, and the model complied with zero harmful single-turn requests about planning a cyberattack across 30 public jailbreak techniques.

For a non-technical business reader, the practical takeaway is that the most capable version stays gated. If your work touches sensitive areas, the public Fable 5 model will sometimes hand off to the older Opus 4.8, which trades some capability for caution.

The pricing cut, less than half the previous Mythos Preview rate, signals Anthropic pushing frontier capability toward wider commercial use while keeping the highest-risk applications behind a trust program.

Frequently Asked Questions

What is the difference between Claude Fable 5 and Claude Mythos 5?

They are the same underlying model. Fable 5 is the public version with safeguards that reroute risky queries to an older model, while Mythos 5 has those safeguards lifted in some areas and is available only to trusted partners such as cyberdefenders.

How much do Claude Fable 5 and Mythos 5 cost?

Both are priced at $10 per million input tokens and $50 per million output tokens. Anthropic says that is less than half the price of its earlier Claude Mythos Preview.

What is Project Glasswing?

Project Glasswing is a collaboration between Anthropic and the US government. Mythos 5 deploys first through it as an upgrade to Claude Mythos Preview, giving cyberdefenders and infrastructure providers access to the model's cybersecurity capabilities.

How often do the safeguards block normal requests?

Anthropic tuned the safeguards conservatively, so they sometimes catch harmless requests. On average they trigger in fewer than 5% of sessions, and risky queries get routed to the older Claude Opus 4.8 model instead.

What benchmark results did Fable 5 achieve?

Anthropic reports Fable 5 scored highest among frontier models on Cognition's FrontierCode coding test, posted the top score on Hebbia's finance benchmark, and broke 90% on its own core analytics benchmark, a 10-point gain over the Opus model.

Anthropic's split release pairs a publicly available Fable 5 with a tightly controlled Mythos 5, advancing frontier AI capability while keeping the riskiest cybersecurity and biology applications behind a trusted-access program. The roughly halved pricing and named enterprise results signal a push toward wider commercial adoption.