On September 1, 2026, Anthropic launched two distinct products based on the exact same underlying model. The public-facing Claude Fable 5.1 sees a price reduction of approximately 25% under typical workloads, while the unthrottled Mythos 5.1 comes with government clearance restrictions. Even with money in hand, ordinary developers cannot buy the world’s top scientific reasoning model.
Scientific Benchmark Scores Double Overnight
Anthropic has released its latest advancements in long-horizon complex task handling. In the Terminal-Bench-Science benchmark, Fable 5.1 recorded a score of 52.6%, more than doubling the previous generation’s performance. On Humanity’s Last Exam (with tools), it reached 65.0%, and achieved 73.4% on CursorBench. The model’s ability to follow through on long-horizon tasks has crossed the critical usability threshold.
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Scientific Research (Terminal-Bench-Science) | 52.6% | 24.7% | 29.0% | 22.4% |
| Coding Intelligence (Terminal-Bench 4.0) | 55.8% | 42.0% | 52.3% | 37.3% |
| General Knowledge (Humanity’s Last Exam) | 65.0% | 63.8% | 63.6% | — |
| Office Automation (AutomationBench) | 31.4% | 17.1% | 26.9% | 19.6% |
While the previous generation scored a modest 24.7% on scientific tasks, the new release more than doubled that result. The most dramatic leap occurred in office automation: jumping from 17.1% in the prior generation to 31.4%, nearly doubling its performance.
Observations from Craig Falls, Head of Quantitative Research at Jane Street, validate these benchmark figures. He noted that Fable 5.1 consistently maintains clean, readable code across multi-step tasks. Long-horizon agentic workflows that used to derail into dead ends after a few hours now stay on track, demonstrating frontier-level trading intuition.
Figure: Anthropic release key visual. Source: Anthropic
Long-Horizon Agents Cut Cost Overhead by 45%
A stronger model has become noticeably cheaper on the open market. Fable 5.1 drops per-token pricing by 25% under typical workloads, and achieves up to a 45% cost reduction in agentic workflows that rely heavily on prompt caching. Base pricing remains at $10 for input and $50 for output per million tokens; the cost savings are achieved entirely through optimized internal cache scheduling. Cache retrieval efficiency now dictates the commercial viability of long-horizon agents.
Fintech platform Ramp deployed Fable 5.1 to run an unsupervised 38-hour machine learning task. The model diagnosed previous anomalies as labeling artifacts, self-corrected the pipeline, and initiated six parallel experiments overnight to complete the job. This dramatic drop in inference costs turns multi-day autonomous trial-and-error into a standard engineering practice.
Figure: Fable 5.1 performance comparison chart. Source: Anthropic
Dissecting Low-Level Libraries to Uncover Legacy Vulnerabilities
False positive rates for safety guardrails dropped by 60% in this generation. Fable 5.1 is explicitly permitted to identify software vulnerabilities, provided it refrains from authoring exploits. Relaxing the static scanning red lines has unleashed the model’s reverse-engineering capabilities.
Investment firm Millennium had a rare system crash that remained unsolved for years. By disassembling third-party libraries and cross-referencing core dump files, Fable 5.1 directly pinpointed the root cause of the memory leak. The model is no longer a text-bound toy restricted to high-level wrappers; it seamlessly probes down to the binary instruction set.
Combined with the Enterprise Frontier Safety (EFS) system launching this fall, data will reside within customer-exclusive cloud infrastructure, enforcing zero-data-retention at the physical layer. Financial institutions can now confidently feed core low-level code directly to the model.
Clearance Access Locks Down Peak Compute Capabilities
The most significant shift lies in how the capability is delivered. Mythos 5.1 and Fable 5.1 share the same base model architecture; their sole distinction lies in guardrail levels. Designed specifically for cybersecurity and life sciences, Mythos 5.1 offers no public API, and access to its biological capability modules is co-developed with the U.S. government.
A registration system for scientists will open soon, but the approval framework establishes clear capability boundaries. On the Terminal-Bench 4.0 test, Mythos reached 60.9%, outperforming Fable’s 55.8%. That extra 5% of peak performance is strictly sealed behind the walled garden of government-cleared partnerships.
The debut of Claude 5.1 marks a definitive watershed moment. Top-tier frontier AI is no longer a off-the-shelf commodity available to anyone with a credit card; it has transformed into a privileged asset requiring clearance and vetting. Ordinary developers can access increasingly affordable lightweight versions, but the elevator to true top-tier capability has been quietly pulled away.
Reference Links:
- Anthropic Official Announcement
- HN Discussion
