Earlier today (US time, September 1),Anthropic Launching Claude simultaneously. Fable 5.1 With Claude Mythos 5.1, the two models share the same underlying architecture, differing only in the level of safety guardrails. In its official announcement, Anthropic called it “the world’s most advanced model for coding and knowledge work,” and emphasized that the focus of this update is not simply on pushing up benchmark scores, but on directly responding to customers’ long-standing criticisms over pricing, data retention, and overly restrictive safety policies.
Fable 5.1 and Mythos 5.1: Same Model, Two Sets of Safety Gates
The Claude family product ladder has always been Haiku (small) → Sonnet (medium) → Opus (large) → Mythos-class (flagship). The Fable 5 series has been Anthropic’s public flagship since its launch on June 9, and version 5.1 builds on that foundation with two improvements: sharpening the precision of the safety filters and lowering operating costs.
Fable 5.1 is now open to all users, available for Claude.ai’s Max/Team/Enterprise plans and API endpoints; Mythos 5.1, however, is only provided to vetted U.S. organizations through Project Glasswing, with looser safety guardrails, suited for scenarios such as cybersecurity defense and life sciences research that require greater operational latitude. Anthropic stated that there is no timeline for expanding access to Mythos 5.1.
Fixed pricing but lower actual costs: the 75% cache price cut is the most noticeable.
The biggest highlight of this update isn’t the listed price: the published rates of $10 per million input tokens and $50 per million output tokens remain unchanged. The real focus is the actual operating cost. Anthropic says Fable 5.1 costs about 25% less than Fable 5 under typical workloads, thanks to a reduction in cache read prices.
When a model re-reads already-processed input (such as a large codebase in a long conversation), this “cache read” was originally billed as a separate line item. This time, Anthropic has slashed the cache read rate for Fable 5.1 by 75%, which is especially significant for AI agent tasks that require long-running operation (coding agents, research agents). According to analysis, Fable 5.1 can reduce costs on complex agent tasks by up to 45%.
Benchmark tests doubled across the board.
Fable 5.1 shows significant improvements across multiple benchmark tests:
- Terminal-Bench-Science 0.1(Scientific Research CLI Test): Fable 5.1 achieved 52.6%, far higher than Fable 5’s 24.7%.
- Terminal-Bench 4.0(Coding test): Fable 5.1 55.8%, Fable 5 42.0%, Mythos 5.1 60.9%, OpenAI GPT-5.6 Sol at 37.3%
- AutomationBenchBusiness process automation: 31.4% vs. 17.1%
- CursorBench 3.0Fable 5.1 outperforms Fable 5, achieving the same results as Fable 5’s highest cost tier within a lower cost range.
- Humanity’s Last Exam: At every effort level, Fable 5.1 achieves the same or better pass rate as Fable 5 at a lower cost.
Anthropic emphasizes that, from the cost-effectiveness curve perspective, Fable 5.1’s competitiveness is more noteworthy than the score itself: to achieve the same score level, Fable 5.1 costs far less than Fable 5.
Safer and more precise: protection trigger rate reduced by 60%
Shortly after Fable 5’s release, Amazon researchers discovered a path that bypassed its cybersecurity filter, allowing the model to successfully identify software vulnerabilities. Following this incident, the U.S. Department of Commerce ordered Anthropic on June 12 to block model access for foreign nationals. The event prompted Anthropic to rethink the design logic of its safety filters, shifting from “blanket blocking” to “targeted interception.”
Fable 5.1 has redesigned its security filtering logic to address this issue. Anthropic says the new cybersecurity filter triggers about 60% less frequently in Claude Code. Fable 5.1 can now perform vulnerability discovery (defensive security work) but will not write exploits for it. Penetration testing work remains reserved for Mythos 5.1, available only to vetted cybersecurity professionals. Anthropic commissioned two external security firms and testing agency Gray Swan to attack-test the new filters, and no major jailbreak vulnerabilities were found.
Enterprise Frontier Safeguards: Zero Data Retention Solution
Along with the model update, Anthropic has also launched the Enterprise Frontier Safeguards (EFS) program. This solution provides zero data retention (ZDR)-level privacy protection on the enterprise side, meaning the model does not store customer data at all, while still maintaining state-of-the-art capabilities to prevent malicious use. For highly regulated industries such as finance, healthcare, and law, this feature may offer more practical value than improved benchmark scores.
EFS will be rolled out to enterprise customers in phases later this year; until then, eligible customers can still use Fable 5. Anthropic emphasized that this solution was developed in collaboration with several major enterprise customers, directly responding to the earlier strong backlash from enterprise users over data retention policies.
Distillation Protection Upgrade: New API Accounts Cannot Edit Historical Messages
To prevent competitors from replicating Fable 5.1’s capabilities through distillation, Anthropic has changed its API policy: newly registered API accounts will not be able to edit Claude’s past response messages, while underlying inference data will be retained. This measure is intended to make model theft more difficult.
Knowledge base updated to June 2026.
Fable 5.1’s knowledge cutoff advances from January 2026 in Fable 5 to June 2026, moving forward by 5 months. The million-token context window and 128,000-token output limit remain unchanged.
Real feedback from 22 early adopters
Anthropic provided testimonials from 22 early customers. Among them, a senior portfolio manager at Millennium noted that Fable 5.1 successfully diagnosed a one-in-a-million system crash caused by a third-party library bug. This issue had existed for several years before being discovered. Jane Street stated that Fable 5.1 resolved more code problems than Fable 5 or Opus 5, and achieved state-of-the-art performance in trading intuition tests. Companies such as Canva, Block, MongoDB, Red Hat, Rakuten, Plaid, and Glean also provided positive feedback, ranging from knowledge Q&A to research drafting, with reviewers preferring Fable 5.1 over Fable 5 at a ratio of approximately 2 to 1.
Platform availability and deployment
Fable 5.1 launched simultaneously on AWS (Amazon Bedrock), Google Cloud Vertex AI, and Microsoft Foundry on its release day. The model is also now available to Max/Team/Enterprise plan users in Claude.ai and the Claude app.
This update clearly shows that Anthropic is shifting from a “pure benchmark-chasing” approach to a practical competitive strategy centered on “cost-effectiveness.” The benchmark results for Fable 5.1 do show significant progress, but the 75% price cut on cache reads and the 45% reduction in agent task costs are the more impactful variables. These figures directly influence whether developers are willing to integrate Claude into long-running production workflows. With pressure from OpenAI’s upcoming next-generation models, Anthropic is leveraging “better and cheaper” as its bargaining chip, targeting enterprise users who prioritize practical value over simply chasing the highest scores.
Source: KOCPC Chinese