Anthropic Ships Claude Fable 5 — First Mythos-Class Model Now Public
Anthropic Ships Claude Fable 5 — First Mythos-Class Model Now Public
Anthropic released Claude Fable 5 today, the first publicly available model in its new Mythos-class tier — the rung above Opus. This is the first model in the Claude 5 family and represents a genuine generational jump rather than an incremental point release. It is available immediately through the Claude API and usage-based Enterprise plans.
The headline benchmark numbers are strong, but the defining feature of this release is not raw capability. Fable 5 is the first major frontier model where the primary engineering achievement is what the model refuses to do.
What Shipped
Fable 5 is the same underlying Mythos brain that Anthropic previously deemed too dangerous to release publicly — the one that triggered a global cybersecurity concern because of its ability to find and exploit software vulnerabilities. Anthropic's solution was not to reduce capability. Instead, they built a public version that hard-blocks high-risk domains and routes those requests to the older, safer Opus 4.8.
The capability race and the safety story are now the same product.
Benchmark Performance
| 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 | 29.3% | 13.4% | 5.7% | — |
The FrontierCode result is the standout. Fable 5 more than doubles Opus 4.8 and laps GPT-5.5 on this brutal production-grade coding benchmark. Anthropic states that the gap widens the longer and more complex the task, which is consistent with a model designed for extended autonomous operation.
Safety Architecture
Fable 5 has surgical guardrails applied in four domains:
- Cybersecurity
- Biology
- Chemistry
- Model distillation
Any request in these areas is refused by Fable 5 and deferred to Claude Opus 4.8. Anthropic's red team spent over 1,000 hours attempting to find a working jailbreak around these blocks and reported zero successes.
This approach represents a shift in how frontier safety is productized. Rather than policy promises or terms of service restrictions, the limitations are hard-coded into the model's inference path.
Pricing
| Detail | Value |
|---|---|
| Input tokens | ~$10 per million |
| Output tokens | ~$50 per million |
| Price vs Opus 4.8 | ~2x |
| Free window | Through June 22, 2026 |
| After June 22 | Usage credits required |
Fable 5 is bundled at no extra cost into Pro, Max, Team, and seat-based Enterprise plans until June 22. After that date, users must burn usage credits to continue using it. Anthropic plans to fold it back into standard subscriptions "as soon as possible."
Mythos 5 Released to Partners
Alongside Fable 5, Anthropic quietly released Claude Mythos 5 — the uncuffed sibling without safety guardrails — to trusted partners only. Early reports indicate strong results in drug design and autonomous genomics research. The gap between what Fable 5 will not do and what Mythos 5 can do is the defining story of this generation.
What This Means
For working developers, Fable 5 is a clear upgrade for hard, long-horizon coding tasks. The FrontierCode and SWE-bench Pro gaps are large enough to be felt on real refactors and migrations. However, any task that brushes against security or biology work will be bounced to Opus 4.8, which matters if that is the domain you operate in.
For teams modeling cost, the 2x price is the catch. The free window through June 22 should be used to benchmark Fable 5 against actual workloads before committing to paid usage. The capability jump may pay for itself on the hardest tasks, but it is unlikely to be worth double on routine work where Opus 4.8 already clears the bar.
The broader signal is that the next axis of competition in frontier AI is not just raw capability — it is deployable capability: power that has been made safe enough to sell at scale. Anthropic just proved its edge is being the company that can release the dangerous thing at all.