Claude 5.1
Version 5.1 of Anthropic's Claude large language model series, featuring strong capabilities in code generation and spatial planning reasoning.
Core Facts
Timeline (last 90 days)
Claude Opus 5.5 performance matches Claude 5.1 on most tasks, input and output pricing have both been reduced, and usage limits for Pro Max and Team users have been increased.
In the multi-robot warehouse scheduling case, Claude 5.1 and GPT-6's path planning implementations are clearer than V4.1 Flash's.
In the online Excel application case, V4.1 Flash's implementation of cell linkage was broadly comparable to Claude 5.1.
The UP host's self-built benchmark has been updated with scores for the Claude series, GPT-5.6, and Claude 5.1, and plans to incorporate DeepSeek's new models and GPT-6.
Claude 5.1's built-in tools increased from 30 to 46.
After using Claude 5.1, overall costs for general tasks decreased by approximately 25%, with savings of up to 45% for heavy Agent tasks.
The complete system prompt of the public version of Claude 5.1 was extracted and leaked by hackers, with the content reaching 275,000 words.
Claude 5.1's score on the commercial automation benchmark improved from 17.1% to 31.4%.
Domestic models such as DeepSeek and Kimi have come close to Claude in everyday conversation, Chinese writing, and general coding scenarios, but still lag significantly behind in Agent capabilities such as long-cycle scientific research tasks, complex reasoning, and deep multi-tool coordination.
Protein binders designed by the trusted version of Claude 5.1 achieved a hit rate of nearly 50% across 12 targets, with binding affinity reaching 10 times the magnitude of the best competition designs on 3 targets.
23 more timeline events
All Facts (20)
Anthropic released Claude 5.1, along with a Claude Opus-level security version for trusted professional users.
65%UnverifiedAfter using Claude 5.1, overall costs for general tasks decreased by approximately 25%, with savings of up to 45% for heavy Agent tasks.
60%UnverifiedClaude 5.1 scored 52.6% on Terminal-Bench Science, more than doubling the previous generation's score of 24.7%.
60%UnverifiedIn the online Excel application case, V4.1 Flash's implementation of cell linkage was broadly comparable to Claude 5.1.
50%UnverifiedIn the multi-robot warehouse scheduling case, Claude 5.1 and GPT-6's path planning implementations are clearer than V4.1 Flash's.
50%UnverifiedClaude 5.1's built-in tools increased from 30 to 46.
50%UnverifiedClaude 5.1's score on the commercial automation benchmark improved from 17.1% to 31.4%.
50%UnverifiedDomestic models such as DeepSeek and Kimi have come close to Claude in everyday conversation, Chinese writing, and general coding scenarios, but still lag significantly behind in Agent capabilities such as long-cycle scientific research tasks, complex reasoning, and deep multi-tool coordination.
50%UnverifiedClaude 5.1 advanced the surface detail of one-third of Venus to 2–3 kilometers using digital elevation models generated from NASA Magellan radar imagery, improving elevation accuracy by approximately 25%.
50%UnverifiedClaude 5.1 wrote custom GPU kernels for 7 open-source protein and genomics models, achieving up to 2.5x speedup per inference with consistent outputs, and reducing GPU costs for genomic analysis by 30–60%.
50%UnverifiedClaude 5.1 can help identify software vulnerabilities in source code but still blocks the generation of exploit code, and binary vulnerability scanning and penetration testing requests are redirected.
50%UnverifiedEvaluation results show that while the trusted version of Claude 5.1 is stronger than the previous generation, it still has not entered the next tier of chemical and biological risk level.
50%UnverifiedClaude 5.1's cybersecurity guardrail false-positive interventions were reduced by an average of approximately 60%, and benign false blocks on basic biology and medical questions decreased by 85% compared to the previous generation.
50%UnverifiedThe Claude 5.1 for the general public and the security version for trusted users are essentially the same underlying model, differing only in the risk boundaries of what they are permitted to handle.
50%UnverifiedThe Claude Code terminal defaults to the High reasoning tier, while the web interface and collaborative scenarios default to the Medium reasoning tier.
50%UnverifiedClaude 5.1 reduces costs for typical workloads by 25%, with reductions of up to approximately 45% for high-agent workloads.
50%UnverifiedOn Terminal-Bench 4.0, the standard version of Claude 5.1 scored 55.8% and the trusted version reached 60.9%; the OS World 2.0 partial completion rate was 77.9%.
50%UnverifiedClaude 5.1 scored 60.9% on Humanities Last Exam without tools, rising to 65% with tools enabled.
50%UnverifiedClaude 5.1 scored 1853 on GDPVal AAV2, surpassing the previous generation's 1723 and also slightly higher than Opus 5.
50%UnverifiedProtein binders designed by the trusted version of Claude 5.1 achieved a hit rate of nearly 50% across 12 targets, with binding affinity reaching 10 times the magnitude of the best competition designs on 3 targets.
50%