Claude Fable 5 reviewed: strong benchmarks, but safeguards limit real-world gains

Anthropic launched Claude Fable 5 as the general-release version of its latest model and Claude Mythos 5 as a less-restricted variant available only to vetted partners starting with Project Glasswing. The two share the same underlying model, with Fable 5 layering on classifiers for cybersecurity, biology, chemistry, and distillation attempts. When those classifiers fire in the Claude app, desktop app, or mobile app, the request is routed to a fallback model — Claude Opus 4.8 at launch — and the user is notified. In the API, flagged requests are blocked by default with a structured refusal reason, though developers can opt into server-side fallback. Anthropic's own system card acknowledges that Fable 5's cyber classifiers consistently fired across tested cyber capability evaluations, meaning Fable's cybersecurity performance is effectively equivalent to Opus 4.8, while Mythos 5 with safeguards off is described as a substantially different story.

The model also performs well across math, long-context reasoning, multimodal tasks, finance, legal work, and tool use.

On benchmarks, Mythos 5 scored 95.5% and Fable 5 scored 95% on SWE-bench Verified, with Mythos at 80.3% and Fable at 80% on SWE-bench Pro. The model also performs well across math, long-context reasoning, multimodal tasks, finance, legal work, and tool use. Pricing is $10 per million input tokens and $50 per million output tokens — less than half the price of Mythos preview according to Anthropic — and Fable 5 is included at no extra cost until June 22nd, after which usage credits will be required unless capacity allows an extension. In AICodeKing's hands-on testing, Fable 5 handled some creative coding tasks well but did not feel like a significant daily-use improvement over Opus or GPT-5.5, and costs can climb quickly for large agentic or benchmark-style prompts. The system card also flags remaining agent concerns including hallucinated testing, prompt injection risks, and evaluation awareness.

Key facts

1. 01Claude Fable 5 (general release) and Claude Mythos 5 (trusted partners only) are the same underlying model.
2. 02Fable 5 adds classifiers for cybersecurity, biology, chemistry, and distillation attempts; triggered requests fall back to Claude Opus 4.8 in consumer apps.
3. 03In the API, flagged requests are blocked by default with a structured refusal reason; developers can opt into server-side fallback.
4. 04Fable 5 scored 95% and Mythos 5 scored 95.5% on SWE-bench Verified; Fable scored 80% and Mythos scored 80.3% on SWE-bench Pro.
5. 05Pricing is $10 per million input tokens and $50 per million output tokens — described by Anthropic as less than half the price of Mythos preview.
6. 06Fable 5 is included at no extra cost until June 22nd; usage credits will be required from June 23rd unless capacity allows an extension.
7. 07In AICodeKing's testing, Fable 5 did not feel like a major daily-use improvement over Opus or GPT-5.5, and costs escalate quickly for large agentic prompts.

Topics