Claude Fable 5: Anthropic's First Public Mythos-Class Model Ships With a Built-In Safety Valve

When Anthropic shipped Claude Fable 5 on June 9, it crossed a threshold no major frontier lab had reached before: a Mythos-class model made publicly available—with capability benchmarks that surpass every prior Claude release and a built-in circuit breaker that silently downgrades to a weaker model when your query touches a restricted domain.

What Changed With the Fifth Generation

Claude Fable 5 scores 80.3% on SWE-Bench Pro, outpacing Opus 4.8’s 69.2% on the same benchmark. It tops Hebbia’s Finance Benchmark for complex analytical reasoning and delivers what The Decoder describes as a new state-of-the-art on vision tasks, including reconstructing functional code from screenshots. Pricing reflects the capability step: $10 per million input tokens and $50 per million output tokens—nearly double what Opus 4.8 costs.

The companion release, Claude Mythos 5, stays restricted to vetted partners through Project Glasswing. It achieves 78% on ExploitBench and reportedly accelerated protein design workflows by 10x in early partner evaluations. Anthropic treats Mythos 5 as too consequential for general availability; Fable 5 is what that capability floor looks like after safety filtering.

The Safeguard That Changes the Architecture Equation

Fable 5 ships with classifiers that automatically route certain query categories—cybersecurity, biology, chemistry, and model distillation—to Opus 4.8 rather than refusing them. Anthropic reports this affects less than 5% of sessions and acknowledges the current filters are “too aggressive,” sometimes flagging harmless content in those domains.

For agentic workflows, that 5% is load-bearing. A pipeline chaining Fable 5 across multiple steps has no deterministic guarantee that every call executes at Mythos-class capability. If a step involves anything adjacent to security analysis, bioinformatics, or competitive intelligence around model architecture, the model steps down silently. As discussed in earlier coverage of agentic AI coding advances, most orchestration frameworks have no visibility into which model tier actually ran on a given call.

Availability rolls out in two stages: Pro, Max, Team, and Enterprise plans get access through June 22, after which usage shifts to a credit-based model rather than unlimited subscription inclusion.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.