Anthropic Debuts Claude Fable 5.1 and Mythos 5.1 With Slashed Agent Costs and Relaxed Safeguards

Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 on Tuesday, releasing its next tier of frontier artificial intelligence models just three months after debuting the original Fable 5 family. The update directly targets developers and enterprise buyers running persistent, long-horizon AI agents by slashing context cache pricing and recalibrating strict safety filters that previously drew customer frustration^[3].

The two releases represent the same core architecture deployed under differing safety perimeters: Fable 5.1 is generally available via the Claude API and cloud platforms, while Mythos 5.1 remains restricted to vetted government and institutional researchers inside Project Glasswing. With competitors like OpenAI previewing powerful new systems, Anthropic is pivoting from pure capability leaps to solving the practical hurdles of cost, over-refusal, and enterprise data privacy.

Lower Caching Costs Target Autonomous Workflows

While Anthropic kept baseline pricing constant at $10 per million input tokens and $50 per million output tokens, it sharply reduced the cost of cache reads by 75 percent, dropping the rate from $1.00 down to $0.25 per million tokens. As reported by The Verge and VentureBeat, the discount significantly alters the economics of persistent agents that must repeatedly re-read expanding codebases or ongoing multi-turn conversations^[3].

According to Anthropic, this change yields an estimated 25 percent overall cost reduction across standard enterprise workloads and can cut expenses by roughly 45 percent for heavy agentic tasks. In developer circles and technical evaluations, the move has been welcomed as an essential adjustment for applications that execute sustained multi-hour sessions.

Dual Release Strategy: Fable 5.1 Versus Mythos 5.1

Anthropic continues to maintain a bifurcated approach to its most potent AI weights. In its official system card, the company noted that Claude Fable 5.1 incorporates consumer-facing guardrails designed to prevent actionable assistance with chemical, biological, radiological, and nuclear threats or malicious cyber exploitation^[9].

Claude Mythos 5.1 unblocks advanced cybersecurity analysis and biological research pathways for authorized partners. In its Responsible Scaling Policy disclosure, Anthropic clarified that the model demonstrates low alignment risk and CB-1 biological capability, falling short of catastrophic thresholds but warranting controlled distribution.

Claude Fable 5.1 and Claude Mythos 5.1 are the same model, but with different levels of safeguards. Fable 5.1 is generally available, while Mythos 5.1 is available only through our trusted access programs.

Anthropic, Official Launch Announcement

Safeguard Refinements and New Enterprise Privacy

Addressing persistent complaints that its models frequently hallucinated safety violations during benign coding, Anthropic overhauled its classifiers. TechCrunch reported that the new safeguards generate roughly 60 percent fewer false positives during typical coding workflows. Fable 5.1 is now explicitly permitted to conduct vulnerability discovery in raw source code, though binary exploit generation remains restricted to Mythos^[9].

To ease corporate governance concerns, Anthropic also announced Enterprise Frontier Safeguards (EFS), developed in collaboration with Amazon Web Services. As detailed in documentation for Amazon Bedrock and Claude Platform on AWS, the architecture allows enterprises to store telemetry and evaluation logs entirely within customer-controlled cloud partitions, matching zero-data-retention standards while keeping abuse monitoring operational.

Performance Benchmarks and Developer Reaction

On technical benchmarks, Anthropic reported major gains in sustained scientific reasoning. Fable 5.1 reached 52.6 percent on the newly published Terminal-Bench-Science benchmark, more than doubling the 24.7 percent score recorded by Fable 5. On Terminal-Bench 4.0, Fable 5.1 achieved 55.8 percent, while the unrestrained Mythos 5.1 scored 60.9 percent, illustrating the exact performance tax imposed by public safety filters^[1].

Initial reaction from developers has been largely positive regarding the API pricing relief, though independent testing communities have urged cautious evaluation. Programmers on developer forums noted that short, non-cached queries will notice negligible savings, making the cost cut relevant strictly to autonomous loops. Furthermore, Anthropic confirmed that Fable 5.1 output now includes content provenance watermarking by default, a measure aimed at synthetic text detection that has already sparked debate over automated code attribution.