Unverified50% confidenceBenchmarkExact time
On the TasteVal benchmark, Opus 5.5 performed at roughly twice the human expert baseline, but task verification is fast and cheap, the human baseline does not include top-tier researchers, and the experiment scale was only 40 GPU hours and a single H100.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/12/2026
First Seen
Valid until: 10/26/2026
Sources
Related Entities
Related Claims
UnverifiedIn a test checking whether AI-modified code could be accepted by a project, Opus 5.5 using the default setting scored slightly higher than Astra's best publicly available score, and the cost per task was approximately one-fifth of Astra's.69% similarUnverified在业务自动化评测中，相同单任务成本下Opus 5的通过率约为次优模型的1.5倍67% similarUnverifiedIn an unknown-rules mini-game test designed specifically for models, GPT-6 scored 99 points, approaching a perfect score, while the highest human score was only in the 30s and other models scored around 40.67% similarUnverified2024年一项研究对GPT-4在Temperature=0下的代码生成任务进行100次测试，发现约8%的测试用例存在跨次运行的输出差异67% similarUnverified即便是GPT-4级别的Agent，面对含有5步以上工具调用的复杂任务时，端到端成功率往往低于60%67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1012993API
curl https://kongchang.com/api/v1/knowledge/claims/1012993MCP
get_claim(id=1012993)