Anthropic's Fable 5.2 costs 2.6× Opus 5.5 per task—and Anthropic says start with Opus
demian_ai · x · 2026-11-10
Anthropic shipped Fable 5.2 but its own launch post tells most people to start with Opus 5.5, its "strongest daily driver." Here's why:
Everyday agent work: Opus is basically there
- OSWorld: Fable 89.3% vs Opus 87.2% (+2 points)
- Terminal-Bench 4.0 (Artificial Analysis): both 60%
- Mergeable code (FrontierCode): Fable high-effort ties Opus's best at roughly half the cost per task
- Max-effort cost per task: Fable $26.5 vs Opus $10 (2.6×)
Long, open-ended jobs: Fable pulls away
- Fermat's Last Theorem formalization with 1,000 agents: Fable finished in 30 hours; Opus stalled at 27%
- Remote Labor Index: Fable 41.3% vs Opus 21.3% and GPT-6 Astra 20.8%
- Two-arm robot sim (RoboDojo): Fable 64.9% vs Opus 39.4%
- One engineer ran 1,000+ subagents for 8 days; it built an OS that boots on a laptop
Take: use Opus for short jobs; Fable earns its price on multi-hour tasks that must stay on track, or as the orchestrator handing easy parts to cheaper models. Notably this launch lands as Anthropic begins IPO marketing Tuesday. Author wants independent long-horizon tests to reproduce the Fermat/freelance gaps.
More from Companies & People
- Cisco enlists Claude Mythos to red-team enterprise AI agents — shashib · 2026-11-10
- Robotics investor heads to CoRL, recruiting for post-training and fleet deployment projects — oyhsu · 2026-11-10
- The race is on to sue Nvidia after board probes Jensen Huang's ties to suspected Chinese spy — TechCrunch AI · 2026-11-10
- Accusation that Jensen Huang misled employees and investors draws calls for his resignation — BlackHC · 2026-11-10
- Atlassian report: 80% of teams ship faster with AI, but customers see no value sooner — aakashgupta · 2026-11-10
- Kai-Fu Lee tells Columbia Business School top AI agents now do a workday in one coffee break — kaifulee · 2026-11-10