AI giant Anthropic has officially launched its highly anticipated fifth-generation Claude series, two new models: Claude Fable5 for the general market and Claude Mythos5, which focuses on specific professional fields (now closed to preview). Both models are built on the same base model but have different emphases on security configurations and application scenarios.
Programming and General Knowledge: Fable5's Dominant Performance
As a general-purpose model, Claude Fable5 achieved the highest scores in almost all major benchmark tests. It particularly excels in long-term and complex tasks:
Software Engineering Breakthroughs: In the SWE-Bench Pro test, which evaluates solving real GitHub tasks without assistance, Fable5 scored an impressive 80.3%, far exceeding Claude Opus4.8 (69.2%) and GPT5.5 (58.6%). On the more stringent production-level coding benchmark FrontierCode, it scored 29.3%, leaving GPT5.5 (5.7%) far behind.
Remarkable Practical Efficiency: Payment giant Stripe stated that Fable5 reduced a project that originally took five months to just a few days; in a codebase with 50 million lines of Ruby code, it completed the entire team's migration work that would have taken over two months in just one day.
Knowledge Work and Visual Leap: Fable5 performed exceptionally well in financial analysis (Hebbia benchmark) and chart interpretation. IMC Trading Group stated that the model nearly passed all of its trading analysis assessments. Visually, it can accurately extract data from complex scientific illustrations and independently complete the game "Pokémon FireRed" based solely on a game screenshot, without relying on auxiliary frameworks required by previous models.
Scientific Hypotheses and Cybersecurity: Mythos5's Autonomous Research
Unlike Fable5, which is equipped with conservative security measures, Claude Mythos5 removes restrictions in areas such as cybersecurity and is specifically available to certain partners and the U.S. government (via the Project Glasswing initiative):
Drug Design Speed Increased by 10 Times: In unaided blind tests, Mythos5 can autonomously select binding sites, run bioinformatics tools, and self-correct errors. Among 14 protein targets, it successfully generated effective candidate drugs for 9 of them.
The First LLM to Propose Scientific Hypotheses: Blind comparisons showed that scientists preferred Mythos5's molecular biology hypotheses in about 80% of cases (for example, a new mechanism for E. coli proteins has been independently verified).
Autonomous Genomic Research: Mythos5 worked continuously for over a week without human intervention, compiling single-cell data from 138 animal species and millions of cells, and training its own machine learning model. Its performance surpassed a model published in Science magazine and was 100 times smaller in size.
Defending Cybersecurity: In the ExploitBench benchmark test, Mythos5's score rose from 69% in the preview version to 78% (Opus4.8 only reached 40%), earning it the title of "the world's strongest cybersecurity model."
A Double-Edged Sword: High Cost and Extreme Security Measures
Along with its powerful performance comes a sharp increase in cost. The pricing for Fable5 and Mythos5 is $10 per million input tokens (MTok) and $50 per million output tokens, nearly twice that of Claude Opus4.8. In the subscription plans on Claude.ai, the new models will be billed at double the usage rate.
To control potential risks such as cyberattacks or bioweapons from Mythos-level models, Anthropic has integrated an innovative classifier degradation mechanism into Fable5:
If the system detects dangerous trigger words related to cybersecurity, biology, chemistry, or "refinement (model capability extraction)," it automatically routes the request to the weaker Claude Opus4.8 model (affecting less than 5% of sessions) and notifies users on the interface.
For prompts aimed at building cutting-edge large models (such as pre-training processes or distributed training designs), the system does not directly block them but subtly limits their output effects through prompt modifications, directional vectors, or PEFT (parameter-efficient fine-tuning).
In external testing lasting over 1,000 hours, testers were unable to find a general jailbreak method, and the success rate of Fable5 attack tasks was zero. To address this, Anthropic has also added a 30-day data retention period to detect new types of attacks.
Launch Timeline
Currently, Claude Fable5 is available through the
