Anthropic has announced two new large language models: Claude Mythos 5, a high-capability model being distributed to a limited set of partners, and Claude Fable 5, a more restricted version intended for general commercial use. The company says the models set new benchmarks on multiple tasks, including software engineering and knowledge work.
What’s new
-
Two models: Claude Mythos 5 and Claude Fable 5. Anthropic has not disclosed details such as architecture, parameter counts, training data, or training methods. Claude Mythos 5 is fine‑tuned for alignment but was not designed to be "safe for general use," while Claude Fable 5 applies additional protective layers.
-
Restrictions and controls: Claude Fable 5 will not provide substantive responses to requests related to cybersecurity, biology, chemistry, distillation, or building cutting‑edge AI. For such prompts the system can refuse to answer or route the request to the less capable Claude Opus 4.8, and it will inform the user when that happens.
Input/output and capabilities
- Inputs: text and images, up to 1 million tokens.
- Outputs: up to 128,000 tokens, with up to 108 seconds to the first token.
- Features: adaptive reasoning (automatically adjusts depth and duration), five levels of reasoning effort (low, medium, high, xhigh, max), tool use, parallel subagents. Claude Fable 5 includes additional safety classifiers.
Performance and independent evaluations
Anthropic stated the models advance the state of the art in several areas. Independent evaluation of Claude Mythos 5 was not available at publication; Anthropic claims Mythos 5’s capabilities match those of Claude Fable 5. Third‑party evaluator Artificial Analysis ranked Claude Fable 5 at the top of its Intelligence Index and multiple component benchmarks:
- With max effort and fallback to Claude Opus 4.8, Claude Fable 5 scored four points ahead of the next best model (Claude Opus 4.8).
- It achieved state‑of‑the‑art metrics on GDPval‑AA (agentic real‑world task performance), Terminal‑Bench Hard (agentic coding and terminal use), 𝜏²‑Bench Telecom (telephone customer service with tool use), AA‑Omniscience Accuracy (factual recall), Humanity’s Last Exam (reasoning based on factual recall), SciCode (scientific coding), and CritPt (physics reasoning).
- Claude Fable 5 topped the AA‑Omniscience Index overall: it led the AA‑Omniscience Accuracy component but ranked 15th on the AA‑Omniscience Non‑Hallucination Rate, indicating it sometimes answers incorrectly instead of refusing or admitting ignorance.
- In domain‑ and language‑specific knowledge assessments, Claude Fable 5 showed the broadest coverage among tested models.
Availability and pricing
- Claude Mythos 5 will be initially provided to selected partners through Project Glasswing.
- Claude Fable 5 is available on consumption‑based pro and enterprise subscription plans; usage credits may apply after June 23.
- API pricing: $10/$50 per 1 million input/output tokens.
- Anthropic will retain "business customer data" for 30 days for the purpose of managing malicious activity and says it will not use those data to train new models.
Safety concerns and limitations
Anthropic rates the models’ propensity to act against a user’s interests as "very low," but it flagged concerns about Claude Mythos 5’s potential to behave undesirably or assist malicious actors. The company warned Mythos 5 could pose a risk to systems that provide "extensive access to sensitive assets" or that have "moderate capacity for autonomous, goal‑directed operation and subterfuge."
Anthropic also said the model is not a substitute for human expertise in activities like developing chemical or biological weapons, but acknowledged uncertainty about whether someone with undergraduate technical knowledge might exploit Mythos 5 for harmful purposes. The company cautioned that Fable 5’s safety measures are imperfect and can in some cases unnecessarily degrade performance.
Controversy and response
At launch, Claude Fable 5 initially included an additional limitation: for inputs related to building highly capable AI (for example, designing pretraining pipelines, distributed training infrastructure, or ML accelerators) the model would degrade effectiveness via methods such as prompt modification, steering vectors, or parameter‑efficient fine‑tuning — and it did so without notifying users. That behavior drew sharp criticism from developers and researchers; Dean W. Ball called undisclosed degradation "shockingly hostile," and Robert Scoble noted the unusually strong community backlash.
Anthropic quickly revised course. It replaced the hidden degradation with a policy that such inputs will either be refused or handed off to a less capable model, and users will be notified when that occurs.
Why this matters
Since Anthropic’s April release of Claude Mythos Preview, security teams have been preparing defenses for potential AI‑assisted attacks, and observers have been probing what this new class of models can do. Mythos 5 and Fable 5 represent a meaningful step forward, particularly in AI‑assisted coding. Splitting Mythos into a fully capable model with limited distribution and a guardrailed version for general use is a pragmatic approach while security teams continue their work.
At the same time, limiting a broadly available product’s ability to assist in legitimately competitive or research activities raises questions about openness, fair competition, and users’ rights to use tools for lawful purposes. The community will likely press for more independent testing and greater transparency about model construction and training.
Conclusion
Claude Mythos 5 signals a substantial capability increase, and Claude Fable 5 brings those capabilities to a wider audience with explicit guardrails. Key development details remain undisclosed, and the initial controversy over hidden capability degradation prompted Anthropic to change its approach. Further independent review and transparency will be important as these models see broader use.
