Anthropic Fable 5.1 Performance Review
the gist
Fable 5.1 offers a significant leap in token efficiency and speed, functioning as a high-power delegation model that outperforms previous iterations in coding, data analysis, and writing tasks.
Performance and Efficiency Gains
Fable 5.1 demonstrates a marked improvement in both latency and token consumption compared to Opus 5. In internal agent benchmarks, Fable 5.1 averaged 766 tokens per run with a 22-second latency, whereas Opus 5 averaged nearly 2,000 tokens per run with a 37-second latency. This efficiency allows for long-horizon task delegation, such as building complex, multi-agent desktop applications, at a lower cost and higher speed than previous Anthropic models.
Coding and Knowledge Work Capabilities
The model excels at end-to-end task completion, particularly in complex coding projects and structured data analysis. When tasked with building a computer-use agent, Fable 5.1 successfully managed 40 sub-agents to deliver a functional application. In knowledge work, the model shows improved discernment, successfully identifying and synthesizing insights from quantitative and qualitative survey data without the tendency to produce generic or "slop" outputs. It also demonstrates superior visual formatting in automated slide deck generation, correctly placing design elements like flow arrows that other models frequently misalign.
Writing and Prose Quality
Fable 5.1 marks a return to the literary quality associated with earlier Claude models, effectively correcting the stylistic issues present in Opus 5 and Sonnet 5. It produces text with higher reading ease and lower grade-level complexity, making it a viable competitor to GPT-4o for drafting blog posts and professional content. The model shows a refined ability to identify structural weaknesses in prose, such as paragraphs that fail to pay off a premise, and provides actionable feedback on how to improve flow and logical connectivity between sentences.