Claude Fable 5 and Mythos 5: Everything You Need to Know About Anthropic's Most Powerful AI Models
Anthropic just released Claude Fable 5 and Claude Mythos 5 — the first models from its new Mythos-class tier that sits above Opus. Fable 5 is the public-facing version; Mythos 5 is the same underlying model with cybersecurity and biology safeguards lifted, available only to vetted defenders and researchers. Here's what developers, engineers, and teams need to know.
What Are Fable 5 and Mythos 5?
Claude Fable 5 is the first Mythos-class model cleared for general use. It's available through the Claude API and consumption-based Enterprise plans. Anthropic calls it the most capable model they've ever shipped publicly, and the benchmarks back that up.
Claude Mythos 5 is the same model with guardrails removed in specific areas. It's deployed through Project Glasswing in collaboration with the US government, focused on cyberdefenders and infrastructure providers. It has the strongest cybersecurity capabilities of any model in the world, according to Anthropic's own evaluations.
The pricing for both is $10 per million input tokens and $50 per million output tokens — double Opus 4.8's rate, which itself may serve as a natural limiter on widespread use.
Benchmarks: Where Fable 5 Dominates
Software Engineering
This is where Fable 5 makes its strongest case. Stripe tested it on a 50-million-line Ruby codebase and reported that Fable 5 compressed a migration that would take a team over two months into a single day.
On SWE-Bench Pro, Anthropic's agentic-coding benchmark:
| Model | Score |
|---|---|
| Fable 5 | 80.3% |
| Mythos Preview | 77.8% |
| Opus 4.8 | 69.2% |
| GPT-5.5 | 58.6% |
| Gemini 3.1 Pro | 54.2% |
On Cognition's FrontierCode Diamond evaluation (the hardest split), Fable 5 reaches 29.3% — more than double Opus 4.8's 13.4% and far ahead of GPT-5.5's 5.7%.
Knowledge Work and Vision
On Hebbia's Finance Benchmark for senior-level reasoning, Fable 5 posts the highest score of any model. IMC reported it aced their trading-analysis evaluations across factual lookup, conceptual reasoning, root-cause analysis, and expected-value analysis.
Vision is a leap. Fable 5 can rebuild a web app's source code from screenshots alone. It beat Pokémon FireRed start to finish using only raw game screenshots with no maps or navigation aids — something earlier Claude models couldn't do even with complex helper harnesses.
On GDP.pdf (vision reasoning over rendered documents, no tools):
| Model | Score |
|---|---|
| Fable 5 | 29.8% |
| GPT-5.5 | 24.9% |
| Opus 4.8 | 22.5% |
| Gemini 3.1 Pro | 16.7% |
Memory and Long-Running Tasks
Fable 5 stays focused across millions of tokens. Anthropic tested it on Slay the Spire with persistent file-based memory — performance improved three times more than Opus 4.8, and Fable reached the final act three times as often.
This is what enables the long-horizon autonomy that companies like Rakuten describe: Fable reflects on and validates its own work, making highly autonomous operations possible.
The Safeguard Architecture
Here's the part that's unprecedented. On certain topics — cybersecurity, biology, chemistry, and model distillation — Fable 5's classifiers will silently hand your query to Claude Opus 4.8 instead. You're told when it happens, but the response comes from a weaker model.
The numbers tell the story:
| Metric | Fable 5 | Mythos 5 | Opus 4.8 |
|---|---|---|---|
| SWE-Bench Pro | 80.3% | — | 69.2% |
| BioMysteryBench (hard) | falls back to Opus | 46.1% | 40.0% |
| ExploitBench (cybersecurity) | falls back to Opus | 78.0% | 40.0% |
On cybersecurity queries, Mythos 5 scores nearly double Opus 4.8 (78% vs 40%). That's exactly why Fable 5 falls back — those capabilities in the wrong hands could cause real damage.
Anthropic says the fallback triggers in fewer than 5% of sessions. They ran an external bug bounty with over 1,000 hours of testing and found no universal jailbreaks, then worked with external red-teaming organizations that also failed to find one.
Mandatory 30-Day Data Retention
With this launch comes a new policy: 30-day retention on all Fable 5 and Mythos 5 traffic, even for enterprises that previously had zero-retention agreements. Anthropic says they won't use the data for training — only to defend against attacks and reduce false positives in the classifiers. This could set an industry precedent where access to the most powerful models comes with mandatory data retention.
Access and Pricing
Fable 5 is available now through the Claude API and Enterprise plans:
- Through June 22: Included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost
- From June 23: Removed from subscription plans, requiring usage credits
- Pricing: $10/million input tokens, $50/million output tokens
- Anthropic expects demand to be very high and difficult to predict
Mythos 5 remains restricted to vetted organizations through Project Glasswing and a trusted access program.
Scientific Breakthroughs with Mythos 5
The most striking results come from Mythos 5's biology capabilities:
- Drug design: Accelerated aspects of the process by ~10x, choosing binding sites, running design tools, and recovering from failures autonomously. 9 of 14 protein targets yielded strong candidates now under investigation.
- Novel hypotheses: In blinded comparisons, scientists preferred Mythos 5's molecular biology hypotheses 80% of the time over Opus-class models. One hypothesis about an E. coli protein was independently corroborated by another lab.
- Genomics: Trained a model on single-cell data across 138 species that outperformed a recent model published in Science — despite being 100 times smaller.
What This Means for Developers
Fable 5 changes the calculus for teams building AI-powered tools:
- Coding agents just got a major upgrade. If you're building agentic coding tools, Fable 5's SWE-Bench Pro lead is significant. The Stripe migration story (two months of work in one day) isn't marketing — it's a data point.
- Vision-heavy workflows are now production-viable. Screenshot-to-code, document reasoning, and complex visual tasks that required heavy scaffolding before now work with minimal setup.
- Budget for the fallback. When Fable 5 falls back to Opus 4.8, you're getting a different model's answer. If your workflow touches cybersecurity or biology, plan for that.
- Token costs are real. At $10/$50 per million tokens, Fable 5 is expensive. For long-running autonomous tasks, budget accordingly — one extended session can burn through usage credits fast.
The Bigger Picture
Anthropic released this model days after warning that AI systems may soon achieve recursive self-improvement. They're simultaneously saying "this is the most powerful model we've ever shipped" and "we need coordinated safety brakes on frontier development." The safeguard architecture — silently routing ~5% of queries to a weaker model — is the practical expression of that tension.
For developers, the takeaway is straightforward: Fable 5 is the new frontier model to benchmark against. It's available now, it's expensive, and it's very, very good at coding and vision. Just know that on certain topics, you won't always get Fable 5's answer.