Fable 5.1 leads, but its top setting costs more
Its highest score came at maximum effort, while a lower setting narrowed the cost trade-off and AWS adds a test route.
By Saeed Ezzati7 min read
The audio edition
Listen to this newsletter
0:001:42
Read transcript
The clearest signal today is that AI performance is still arriving with a bill attached. Anthropic’s Fable 5.1 took the top measured spot on Artificial Analysis’s Intelligence Index, scoring 66 at its maximum effort setting—the highest score that evaluator has recorded. But that run was estimated at three dollars and 76 cents per benchmark task, versus three dollars and 14 cents for Fable 5, a 20 percent increase. The crucial caveat is that this is not one fixed model behavior. Across five effort settings, Fable 5.1 scored from 58 to 66, while output-token use ranged from 13.1 million to 143.7 million tokens per task. At the less intensive xhigh setting, it scored 65 for an estimated two dollars and 72 cents. Anthropic kept its input, output, and cache-write prices, while cutting cache-read pricing to 25 cents per million tokens. Artificial Analysis says that saving reduced the task cost by about a dollar and 40 cents, but Fable 5.1 still used roughly 1.7 times as many output tokens. And the lead is not decisive everywhere. Fable 5.1 and Opus 5 had overlapping confidence intervals on GDPval-AA v2, and were effectively tied on AA-Briefcase. About 4 percent of evaluation output tokens came from fallback to other Claude models, so safeguards and routing are part of the observed result too. Fable 5.1 is generally available on AWS, giving enterprises a practical way to test whether the extra reasoning pays off on their own workloads—or simply creates more work and cost downstream. That downstream cost is already visible in freelance work. Freelancer.com says listings for correcting AI output rose 87 percent to 10,760 worldwide between August 2025 and June 2026. The jobs include rebuilding images, repairing robotic voiceovers and malformed 3D models, and revising prose that feels repetitive or emotionally flat. Upwork reported a 70 percent rise in AI-remediation gigs, while Fiverr said searches for AI-cleanup services increased more than twentyfold since 2023. Those are listings and searches, not completed projects or earnings. And the repair is not always quick: illustrator Todd Van Linda said cleanup can take hours or days. The broader pattern is a trade—faster first drafts, followed by human verification and reconstruction. The same verification problem is emerging in hiring. After five AI interviews produced no follow-up, a job seeker identified as Christopher used ChatGPT Voice to impersonate him in a ten-minute call with Riley, an AI recruiter used by Everforth Apex Systems. A fictional résumé-matched applicant also completed a 23-minute interview without a visible next step. That is one reported experiment, and Everforth did not respond to WIRED. It does not establish the company’s criteria or how widespread synthetic applications are. But Greenhouse reported that 63 percent of surveyed U.S. job seekers had faced an AI interview, while 51 percent of candidates who completed one never heard back. Automation is now screening automation, without necessarily improving accountability. And in robotaxis, the useful safeguard is explanation close to the decision itself. MIT and Motional’s CW-Net places readable concepts—such as “close to cyclist”—inside the planner’s final trajectory choice, producing an explanation in real time rather than after the fact. In a private-track test, it revealed that a vehicle had selected a path that would have hit a cyclist; emergency braking prevented the impact. The system was trained on 130 million labeled scenes, and reported driving capability changed by less than 1 percent. Across these stories, the thing to watch is not just whether AI gets better. It is whether effort settings, cleanup labor, feedback loops, and interpretable safeguards make that improvement usable in the real world.
