Anthropic Releases Claude Fable 5 and Mythos 5 With Major Gains in Coding and Science
Anthropic has launched two new AI models — Claude Fable 5 and Claude Mythos 5 — delivering significant improvements in coding, science, and math performance. The models are now available via API and the claude.ai platform.
The release marks a major leap over previous generations. Anthropic reports Fable 5 outperforms its predecessor on multiple benchmarks, especially in programming tasks and complex scientific reasoning.
What the New Models Bring
Both models are designed for enterprise and advanced research users. Key capabilities include:
- Coding proficiency: Fable 5 achieves state-of-the-art scores on coding benchmarks like SWE-bench, HumanEval, and CodeContests. It can generate, debug, and refactor complex code with higher accuracy.
- Scientific reasoning: Mythos 5 excels on graduate-level biology, chemistry, and physics questions. Benchmarks show improved performance on tasks requiring multi-step logical deduction.
- Math performance: Both models show double-digit percentage gains on challenging math datasets, including GSM8K and MATH.
Anthropic said the models use a new training technique that improves the alignment between the model’s internal reasoning and its output, reducing hallucination rates.
Available Now via API and claude.ai
Claude Fable 5 and Mythos 5 are available immediately. Users on the Anthropic API can access them. The claude.ai web interface also supports the new models.
Anthropic did not disclose pricing changes. Existing API pricing tiers remain in effect.
Why This Matters
The improvements address two of the most demanding use cases for large language models: generating reliable code and solving technical scientific problems. Enterprise customers, researchers, and developers are the primary beneficiaries.
Anthropic positions these models as a direct competitor to OpenAI’s GPT-4 and Google’s Gemini families, especially in the enterprise AI market.
Benchmark Highlights
Anthropic released selected benchmark results for both models:
- SWE-bench (coding): Fable 5 scores 48.5%, up from 35.2% on the prior model. This represents a 38% relative improvement.
- MMLU (massive multitask language understanding): Both models exceed 88% accuracy, matching or beating GPT-4 on certain categories.
- GPQA (graduate-level Q&A): Mythos 5 scores 71.2%, a 12 percentage point gain over the previous version.
Anthropic noted that these results are based on internal evaluations and have not been independently verified.
New Safety and Alignment Features
The models incorporate updated safety guardrails. Anthropic added a new layer of output monitoring that detects and blocks potentially harmful code or scientific advice.
The company also trained the models on a larger dataset of vetted technical documentation to reduce the likelihood of generating insecure code or incorrect scientific information.
“Our goal was to make these models not just more capable, but more trustworthy for high-stakes applications,” an Anthropic spokesperson said.
Comparison to Previous Models
Claude Fable 5 replaces Fable 4. Claude Mythos 5 replaces Mythos 4. Both older models remain available for now, but Anthropic recommends users migrate to the new versions.
Key differences:
- Speed: Fable 5 is roughly 20% faster at generating long code outputs.
- Context window: Both models support the same 200,000 token context as before.
- Multilingual ability: Improvements extend to non-English coding tasks and scientific texts.
Availability and Limitations
The models are available in all regions where Anthropic offers service. They support the same API parameters as previous Claude models.
One limitation: Anthropic warned that the models still struggle with highly specialized or niche scientific subfields. Users should verify outputs for critical applications.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.