Claude Fable 5 and Claude Mythos 5: Everything You Need to Know
Anthropic just shipped its most powerful public model ever — and a restricted twin. Benchmarks, pricing, safeguards, real-world results, and what it means for you.
Advertisement
Anthropic released two models today, and the split between them is the real story. Claude Fable 5 is the most capable model the company has ever made generally available — state-of-the-art on nearly every benchmark it published. Claude Mythos 5 is the same underlying model with safeguards lifted in select areas, and you almost certainly cannot use it.
That two-tier release is new for the industry. One frontier model, two doors: a public one with safety classifiers bolted on, and a restricted one reserved for vetted cyberdefenders and biology researchers working with the US government. Whether you find that reassuring or frustrating says a lot about where you sit in the AI ecosystem.
We have read the announcement, the system card coverage, and the early reporting. Here is everything that actually matters — benchmarks, pricing, the safeguard mechanics, what early customers achieved, and the gap between the hype and what you will experience on day one.
What is Claude Fable 5?
Fable 5 is a Mythos-class frontier model tuned for general availability. In plain terms: Anthropic trained a model it considers too capable to release unguarded, then wrapped it in AI classifiers that watch for a narrow set of high-risk requests. Everything else — coding, analysis, writing, vision, agentic work — runs at full strength.
The headline capabilities, per Anthropic: state-of-the-art at coding, knowledge work, vision, and computer use; the first model to break 90% on its core analytics benchmark of complex, long-running analytical tasks; and roughly 25-30% faster task completion than Claude Opus 4.8 on certain benchmarks. It can also work autonomously for longer than any previous Claude model — Anthropic describes agent runs lasting days.
The model ID is claude-fable-5, live on the Claude API today.
What is Claude Mythos 5?
Mythos 5 is the same model without the safety classifiers in specific domains — primarily offensive cybersecurity and advanced biology. It is not for sale. Access runs through Project Glasswing, a collaboration with the US government, plus a planned Trusted Access Program for vetted biology researchers.
The numbers explain why Anthropic kept it gated. On ExploitBench, a cybersecurity benchmark, Mythos 5 scores 78% — against 69% for the earlier Mythos Preview and just 40% for Opus 4.8. A model that nearly doubles the previous public ceiling at finding and exploiting vulnerabilities is exactly the kind of capability you route to defenders before attackers.
Fable 5 vs Mythos 5 at a glance
| Claude Fable 5 | Claude Mythos 5 | |
|---|---|---|
| Underlying model | Identical — same weights, same training | |
| Safeguards | Classifiers active for cyber, bio/chem, distillation | Lifted in select areas |
| Who can use it | Everyone (API, Claude apps) | Project Glasswing partners, vetted researchers |
| Pricing | $10 / $50 per million tokens | Same, where available |
| Fallback behavior | Flagged queries route to Opus 4.8 | No fallback in unlocked domains |
| Data retention | 30-day retention on all Mythos-class traffic (not used for training) | |
The benchmarks
Anthropic published an unusually aggressive comparison set. The coding numbers are the standouts.
Software engineering
| Benchmark | Fable 5 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| SWE-bench Pro | 80.3% | 69.2% | 58.6% | 54.2% |
| FrontierCode (Cognition) | 29.3% | 13.4% | 5.7% | — |
Two things stand out. The 11-point jump over Opus 4.8 on SWE-bench Pro is the largest single-generation gain Anthropic has shipped since the 4.x line began — we tracked that lineage in Claude 4.8 vs 4.7. And FrontierCode, built by Cognition to be brutally hard production-level work, shows Fable 5 more than doubling Opus 4.8 and landing 5x ahead of GPT-5.5. On genuinely difficult coding, this is not a close race right now.
Knowledge work
| Benchmark | Fable 5 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| GDPval-AA (Elo-style) | 1932 | 1890 | 1769 | 1314 |
GDPval-AA measures performance on economically valuable white-collar tasks. The 42-point gap over Opus 4.8 is meaningful but not dramatic — the gap over Gemini 3.1 Pro is. Anthropic also cites the top score of any model on Hebbia's finance benchmark and state-of-the-art results on vision tasks like reading diagrams, charts, and tables nested inside PDFs.
Where Mythos 5 pulls away
| Benchmark | Mythos 5 | Mythos Preview | Opus 4.8 |
|---|---|---|---|
| ExploitBench (cybersecurity) | 78% | 69% | 40% |
This table is the entire argument for the two-tier release in one row. The capability exists; Anthropic is choosing who gets it.
Pricing: double Opus, half of Mythos Preview
| Model | Input / M tokens | Output / M tokens |
|---|---|---|
| Claude Fable 5 | $10 | $50 |
| Claude Opus 4.8 | $5 | $25 |
Fable 5 costs exactly twice Opus 4.8, and less than half of what Claude Mythos Preview cost. The existing 90% prompt-caching discount on input tokens applies, which matters enormously for agentic workloads where the same context gets re-read across many turns. US-only inference is available at a 1.1x multiplier. One subtle detail: when a query trips a safeguard and falls back to Opus 4.8, you are not billed at Fable pricing for that response.
Subscribers get a taste first: Fable 5 is included in Pro, Max, Team, and Enterprise seat plans from June 9 through June 22, after which it requires usage credits. The API and consumption-based Enterprise plans have it immediately, no window.
What early customers actually did with it
The launch results are more interesting than the benchmarks:
- Stripe used Fable 5 to perform a codebase-wide migration on a 50-million-line Ruby codebase in a day — work estimated at over two months for a full team. Anthropic quotes it as compressing five months of engineering into days.
- Drug design: protein design experts using Mythos 5 reported roughly 10x acceleration in parts of the process, with 9 of 14 protein targets yielding strong candidates — the model matching or beating skilled human operators.
- Genomics: a Mythos 5 agent ran largely unattended for over a week on research spanning 138 animal species.
- Long-horizon memory: in Anthropic's Slay the Spire agent testing, giving Fable 5 persistent file-based memory improved its performance three times more than the same setup improved Opus 4.8 — it is meaningfully better at using what it writes down.
Worth noting: every launch-day customer story is a vendor-selected best case. Stripe's migration is real and impressive, but it tells you the ceiling, not the average. Your first week with Fable 5 will look more like "noticeably better Opus" than "two months of work in a day."
How the safeguards actually work
This is the most novel part of the release. Fable 5 ships with AI classifiers monitoring three areas: offensive cybersecurity (exploitation, agentic hacking), biology and chemistry (applied broadly), and attempts to distill the model. When a request trips a classifier, it is silently routed to Claude Opus 4.8, which answers instead.
Anthropic says the system is tuned conservatively — more than 95% of sessions involve no fallback at all — and that over 1,000 hours of external bug-bounty testing produced no universal jailbreaks. The flip side: if you work in security research or the life sciences, you may hit the fallback on legitimate work and get a weaker answer with no recourse beyond the Trusted Access Program.
The other condition of Mythos-class access is a mandatory 30-day data retention policy on all traffic, which Anthropic frames as necessary for detecting novel attacks. The data is not used for training, but if your compliance posture forbids retention, that is a real constraint today.
Expectations vs reality
| The expectation | The reality |
|---|---|
| "Mythos for everyone" | The public gets Fable 5 — Mythos capabilities minus cyber and bio. For most users this distinction will never surface; >95% of sessions see no fallback. |
| "Benchmarks translate directly" | The coding gains are real and you will feel them. The knowledge-work gap over Opus 4.8 (1932 vs 1890) is an increment, not a leap. |
| "Agents that work for days" | True in Anthropic's harnesses with persistent memory. Out of the box, you still need to build the scaffolding — memory files, checkpoints, verification. |
| "Frontier model, frontier price" | 2x Opus 4.8 stings, but prompt caching at 90% off input makes long agentic sessions cheaper than the sticker suggests. |
| "Safety theater" | Partly fair criticism, partly not. TechCrunch notes the release came days after Anthropic warned frontier AI needs brakes. But routing to a weaker model and gating Mythos behind vetting is a concrete mechanism, not just a press line. |
What to expect if you build with it
Based on what is verified so far, here is the practical read:
- Coding agents get the biggest lift. The SWE-bench Pro and FrontierCode deltas are where the money went. If you run Claude Code or agentic IDE workflows, this is the upgrade that justifies the price.
- Long-horizon work is the new frontier. The 3x memory-leverage result means Fable 5 rewards harness design — give it files to write state into and verification loops, and it compounds. Our skills vs agents explainer covers the architecture that benefits.
- Most chat users will notice speed more than smarts. 25-30% faster completions than Opus 4.8 is the change you feel daily; the intelligence delta shows up on hard tasks.
- Budget for 2x or cache aggressively. At $10/$50, naive prompting gets expensive fast. Structured context reuse is no longer optional.
- Security and bio teams should apply for trusted access rather than fighting the classifiers.
How to try it today
API users: model ID claude-fable-5, live now. Claude app subscribers on Pro, Max, Team, or Enterprise: included free through June 22, usage credits after that. Mythos 5: you need to be a Project Glasswing partner or wait for the Trusted Access Program — there is no waitlist link, and Anthropic is explicit that general access is not planned for now.
The verdict
Fable 5 is the strongest public model you can use today, by a margin that is decisive in coding and real but modest elsewhere. The two-tier Fable/Mythos structure is the more consequential precedent: Anthropic has decided that frontier capability and public access are now separate decisions, made per domain. Expect every major lab to copy some version of this within a year. If you build agentic or coding products, upgrade and re-benchmark now — the June 9-22 subscription window makes trying it free. If you mostly chat, Opus 4.8 at half the price remains the value pick, and we compared it to the competition in Claude Opus 4.8 vs GPT-5.5.
Keep going
For every published score in one place — including where GPT-5.5 wins — see Claude Fable 5 vs Opus 4.8 vs GPT-5.5: Benchmarks. For the back-story on the lineage Fable 5 replaces at the top, read Claude 4.8 vs 4.7: What Actually Improved and 10 Things Claude Opus 4.8 Does Better Than GPT-5.5. To see what agentic coding looks like in practice, we let Claude build a full SaaS app. And if you want your AI doing more out of the box, browse installable Agent Skills — they work with Fable 5 from day one.
Frequently Asked Questions
9 questions answered