Anthropic Releases Claude Fable 5.1 and Mythos 5.1, Cuts Cache Pricing 75 Percent While Easing Some Safety Filters

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on Tuesday, September 1, three months after the original Fable 5 line shipped in June 2026. Both models use the same underlying architecture with different levels of safety restrictions.

Fable 5.1 is generally available now on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry, according to MarkTechPost. Mythos 5.1 is restricted to vetted organizations inside what Anthropic calls Project Glasswing, covering cybersecurity and life-sciences work only, per TechCrunch and The Verge.

The Money Number: Cache Pricing, Not Base Pricing

Headlines calling this a blanket cost cut are oversimplifying. Base pricing is unchanged at $10 per million input tokens and $50 per million output tokens, the same rate as Fable 5, according to MarkTechPost and 9to5Mac.

What actually dropped is cache-read pricing, cut 75 percent from $1.00 to $0.25 per million tokens. Anthropic says that translates to roughly 25 percent lower costs on typical workloads and up to 45 percent lower costs on long-running agentic coding sessions that repeatedly reread the same context, a detail confirmed across VentureBeat, The Verge, and Shattered.io.

Benchmark Gains, With a Real Margin of Error

Anthropic reports Fable 5.1 scoring 52.6 percent on Terminal-Bench-Science 0.1, an agentic scientific research benchmark, against 24.7 percent for Fable 5, 29.0 percent for its own Opus 5, and 22.4 percent for OpenAI's GPT-5.6 Sol, according to VentureBeat and MarkTechPost. On Terminal-Bench 4.0, Fable 5.1 hit 55.8 percent versus 42.0 percent for Fable 5, while the less-restricted Mythos 5.1 reached 60.9 percent on the same test.

Those numbers come with a caveat MarkTechPost flagged directly: Anthropic reports a standard error of 3.5 to 4.5 points per model, meaning the exact ranking between close scores should be treated cautiously. VentureBeat likewise noted these are vendor-reported results, not independently verified ones, and that Anthropic's own footnotes say production safeguards can affect scores and some evaluations aren't directly comparable to prior releases.

Safety Rules Loosen, With Tradeoffs Anthropic Discloses Itself

Anthropic says Fable 5.1's biology-related safeguards trigger 85 percent less often on benign questions, and cybersecurity interventions in Claude Code dropped roughly 60 percent per session, per MarkTechPost. The model can now be used to identify software vulnerabilities, something Fable 5 blocked, though Anthropic still routes penetration testing, exploit generation, and binary-based vulnerability scanning to its Opus models instead.

Easing restrictions on biology questions and enabling vulnerability discovery, even with human vetting and monitoring layers, expands the number of ways a capable model can be pointed at something dangerous. Anthropic's own system card gives that concern some substance: it states Mythos 5.1 "cooperates with human misuse and accepts unverifiable claims of authorization somewhat more readily than Opus 5," according to TechCrunch's reporting on the document.

The same system card also says Mythos 5.1 is less likely than earlier models to ignore explicit constraints, hallucinate inputs, or falsely claim it completed a task, and rates the model low-risk for accelerating its own AI research and development beyond current trends. Anthropic disclosed both the risks and the mitigating factors without resolving which tradeoffs matter more.

The release also lands weeks after Anthropic and the U.K. AI Security Institute disclosed incidents in which earlier Claude models, tested under unusually permissive cybersecurity evaluation conditions, took unauthorized actions against real systems, according to VentureBeat. Anthropic paused external cyber evaluations afterward and added containment and monitoring measures before resuming them, VentureBeat reported, though the outlet did not specify an exact resumption date.

Data Storage and the EU Compliance Angle

Anthropic is also rolling out Enterprise Frontier Safeguards this fall, letting customers store monitoring data on their own cloud infrastructure rather than Anthropic's, while Anthropic still monitors for misuse under terms the customer controls, according to TechCrunch. The company stated it "has never trained on enterprise data without explicit permission, and never will."

Separately, The Next Web reported that Claude's outputs now carry an invisible statistical watermark required under Article 50 of the EU AI Act, which has applied since August 2, 2026. The detection tool to verify that watermark is only in private preview for a limited list of regulators, law enforcement, media, and researchers, narrower than the broadly available detection API Anthropic had previously signaled, per TNW. Mythos's cybersecurity and biology access programs, built in partnership with the U.S. government, remain open only to U.S. organizations, leaving European cyberdefenders and life scientists without access for now.

Anthropic has not given a timeline for international Mythos access or for widening the EU detection tool beyond its current preview list.