Robotic arms have a well-documented history of being frustratingly bad at picking things up. Anthropic’s Claude Fable 5.1, released September 1, 2026, just made a significant dent in that problem.
The model achieves a 40% success rate on robotic pick-and-place tasks, up from 5% for its predecessor Claude Fable 5.
What changed under the hood
Fable 5.1 is engineered for long-horizon agentic work, the category of tasks where a model needs to plan, execute, and recover across many steps without a human holding its hand. Coding pipelines, scientific research workflows, and complex business automation are the target terrain.
The benchmark numbers reflect that focus. On Terminal-Bench-Science, Fable 5.1 scores 52.6%, compared to 24.7% for Fable 5. OSWorld 2.0 strict completion climbs from 36.1% to 41.7%. AutomationBench, which tests multi-step software automation, rises from 17.1% to 31.4%.
Customer feedback after launch echoed the benchmark story. Enterprise users specifically cited improved reliability on projects that run for hours, not minutes, with the model completing those tasks using fewer tokens than prior versions.
The pricing math is more interesting than it looks
Base pricing for Fable 5.1 holds steady at $10 per million input tokens and $50 per million output tokens, the same as the previous model.
The cache-read cost dropped 75% to $0.25 per million tokens. Anthropic estimates the cache reduction translates to roughly 25% savings on typical workloads. For more complex agentic pipelines, the savings can reach 45%.
The model is available to Pro, Max, Team, and Enterprise users, and is accessible via the API and major cloud marketplaces.
Why the robotics number deserves its own conversation
The pick-and-place improvement is the headline figure, but it is worth sitting with why that specific benchmark matters. Pick-and-place is a proxy for physical-world grounding: can an AI model translate abstract reasoning into precise, real-world motor commands? The task involves object recognition, spatial reasoning, grasp planning, and error recovery when things go wrong.
A 5% success rate is essentially noise. A 40% rate is not yet good enough for unsupervised industrial deployment, but it crosses the threshold where the technology becomes useful as a human-assist layer rather than a curiosity.
Competitive context
Independent evaluators ranked Claude Fable 5.1 at or above prior leaders on intelligence indices shortly after its release. The pattern of improvements across Terminal-Bench-Science, OSWorld 2.0, and AutomationBench suggests Anthropic is targeting enterprise customers running complex, multi-step, multi-hour workflows rather than consumers asking quick questions.