Claude Fable 5.1 and Mythos 5.1: Same Model, Different Safeguards
Anthropic released Claude Fable 5.1 and Mythos 5.1, sharing one model with different safeguards. Fable is for general users; Mythos is limited to vetted experts, with a 75% cached-input price cut.
Anthropic unveiled two new foundation models on September 2, 2026: Claude Fable 5.1 and Claude Mythos 5.1. They share the same underlying model but apply different protection tiers, so their use cases and audience differ.
What exactly differs between Fable and Mythos
Fable 5.1 is available on every platform. Mythos 5.1 applies stricter safeguards for cybersecurity and biosecurity research, so access is limited to verified individuals and institutions through an approved application process.
Even when the same model is deployed with only safeguards adjusted, operations become more complex. Anthropic says it balances safety, reliability, and utility through a unified approach.
Price cuts that are actually felt
The biggest change in this announcement is a 75% reduction in cached-read pricing. For workloads that reuse many inputs, overall costs drop by up to 45% versus Fable 5, and routine agent tasks become about 25% cheaper.
Pricing per token is 10 tokens per million for input and 50 tokens per million for output, the same as Fable 5. Most savings appear during the cached-read phase.
Benchmark performance: how far it has come
Anthropic evaluated Fable 5.1 on Terminal-Bench-Science 0.1, Terminal-Bench 4.0, Humanity's Last Exam, and CursorBench 3.2.0. Key comparison figures are below.
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 |
|---|---|---|---|
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% |
| Terminal-Bench 4.0 | 55.8%~60.9% | 42.0% | 52.3% |
| Humanity's Last Exam | 65.0% | 63.8% | 63.6% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% |
On Terminal-Bench 4.0, both Fable 5.1 and Mythos 5.1 lead the rankings; Anthropic notes their scores diverge once safeguards are activated.
Solving hard problems: drought, heat, biology
Drought-resistant crop design
After providing drought-resistance design tools for Mythos 5.1, Anthropic tested designs at two external labs. On targets including EGFR, Nipah G, and 15-PGDH, binding affinity reached up to 10 times the previous leaderboard best for the Adaptyv Bio drought-resistance challenge, and validity across 12 targets was about 50%. This is a meaningful improvement over the 10–15% typical in ordinary protein design.
Heat-map generation
Fable 5.1 used NASA Magellan radar imagery and existing heat maps to improve trajectory resolution. The new map covers about one-third of Venus, improving surface resolution from the previous 10–20 km to 2–3 km, and cutting temperature uncertainty by up to 25%. The results were published under Creative Commons and will support NASA VERITAS and ESA EnVision observation planning.
Computational biology acceleration
Mythos 5.1 generated user-specified GPU clusters and cached intermediate results to accelerate up to 2.5 times on 7 open-source biology models. Results were consistent; in analyses where fewer models were run per sequence, GPU costs dropped by 30–60%.
Industry safeguards and privacy
With this announcement, Anthropic also introduced Enterprise Frontier Safeguards (EFS). Even when customers store data in a private cloud for regulated-data protection and similar guarantees, Anthropic maintains that abuse and takeover risks remain low. Financial services, healthcare, pharmaceuticals, telecommunications, legal, and public-sector customers from more than 100 countries participated, along with AWS, Google Cloud, and Microsoft Azure. Supported products include Claude Code, Claude Enterprise, Amazon Bedrock, Google Agent Platform, and Microsoft Foundry.
In biology safeguards, a whitepaper shows a ~85% success rate in blocking requests for normal biology and medical use, while cybersecurity safeguard activation frequency fell by about 60%. Residual-risk tests, red-teaming, jailbreak generation, and backdoor-point inspections were still conducted on the more capable Opus model.
Safety assessment reality and limits
Mythos 5.1 performed strongly on chemical, biological, radiological, and nuclear threat evaluations, but it did not meet the next Responsible Scaling Policy threat tier. It also remains below the higher thresholds in the Frontier Compliance Framework for cyber evaluations.
On most alignment benchmarks, Mythos 5.1 improved over Mythos 5, but deep long-context tasks and multi-agent environments are still difficult to monitor fully, a caveat Anthropic included. Cases that exploit chain-of-thought or automated model decomposition are expected to increase.
For models released after August 2, 2026, EU AI Act compliance requires watermarking outputs to indicate text-generation capability; Anthropic is releasing its detection API under an open preview.
Practical implications: who, and when, should use what
This release centers more on model differentiation and safety-capability separation than raw capability competition. Fable 5.1 delivers better token-reuse efficiency for long-context agent work, cutting ordinary operational costs by 25%.
Enterprise Frontier Safeguards give regulated industries private-cloud data protection and equivalent guarantees, yet insider-threat risks remain possible. Points that became clearer because of safeguard activation include expanded red-team testing, jailbreak analysis, and biosecurity operational research.
Additionally, Mythos 5.1 is currently available only within the United States, and its biosecurity access program is in an early registration state within the U.S. federal contract system. For general users, direct use is limited.
Teams with high security requirements should verify cached-input reuse rates. Cache hit rates below 30% make latency and price improvements of 25% unreliable.
References
- Anthropic announcement: Claude Fable 5.1 and Mythos 5.1 announcement
- GeekNews thread: Claude Fable 5.1 and Mythos 5.1 announcement
- System card: Claude Fable 5.1 and Mythos 5.1 system card
Related posts
67 Cents, 1.5 Hours to Train a Tiny Transformer Beating LLMs on ARC-AGI-1
Mithil Vakde trained a tiny Transformer from scratch on one RTX 5090 for $0.67 and 1.5 hours, scoring 44% on ARC-AGI-1 and 7% on ARC-AGI-2. It beats multiple LLMs and matches TRM/HRM.
Agent Behavior Standard: A New Way to Evaluate AI Agents
Agent Behavior standardizes repeated AI agent actions. This framework makes long-running agent behavior observable and measurable across recognition, judgment, execution, and recovery.
GLM-5.3 openweight shifts coding, agents, and security at once
Z.ai released the GLM-5.3 openweight. Post-training alone lifts coding, agent, and security benchmarks on the same base model, and weights/configs/tokenizers can be deployed or fine-tuned directly.
Curated, fact-checked, and edited by a single operator before publishing.