Anthropic’s Claude Fable 5 completed 16.1% of real freelance projects to a standard judged as good as or better than a paid human professional, according to updated results published this week by the Center for AI Safety, roughly double the previous best score recorded on the same benchmark.
The Remote Labor Index, developed jointly by the Center for AI Safety and Scale AI Labs, tests AI models against 240 real, economically valuable freelance projects spanning categories including 3D and CAD design, architecture, graphic design, video and animation, audio production, data analysis and web application development. Every submission is evaluated by human judges against a gold-standard deliverable produced by a paid professional for the same project.
Fable 5’s 16.1% automation rate nearly doubles Claude Opus 4.8, which scored 8.3% in the same test round, and more than triples OpenAI’s GPT-5.5, which reached 6.3%. All three newly evaluated models scored above every model previously tested on the benchmark. When the Remote Labor Index launched, the best-performing AI agent automated just 2.5% of projects; the previous published leader, Opus 4.6 paired with Anthropic’s Cowork tool, sat at 4.17%.
“The frontier has more than quadrupled in under eight months,” the Center for AI Safety said in its published analysis, calling the jump “a concrete signal of how quickly economically capable AI agents are advancing.”
Researchers were able to evaluate Fable 5 on 218 of the benchmark’s 240 projects before the model’s access was restricted by a U.S. export control directive on June 12. Anthropic restored global access to Fable 5 on July 1 after the directive was lifted, pairing the relaunch with a retrained cybersecurity safety classifier the company says blocks a previously reported jailbreak technique in more than 99% of cases.
The Center for AI Safety also tested an automated “LLM judge” designed to replace human evaluators as the benchmark grows more expensive to score at the frontier. Calibrated against older models, the automated judge significantly overstated performance when applied to Fable 5 and GPT-5.5, inflating scores by roughly 2.5 to 3 times compared with human evaluation, according to the organization, underscoring that human judgment remains necessary to accurately assess the newest models.
Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, twice the cost of Opus 4.8. Researchers noted Fable 5 required a larger per-project budget cap, $150 compared with $50 for other models, to complete a comparable share of tasks without running over cost limits.
