Claude Opus 5 Tops Knowledge-Work Leaderboard, Beating Fable 5 and GPT-5.6

The benchmark result with the broadest economic implications from Claude Opus 5's launch isn't about code: Opus 5 tops GDPval-AA v2, the knowledge-work evaluation, with an Elo of 1,861 — ahead of Claude Fable 5 (1,747) and GPT-5.6 Sol (1,736).

What GDPval Measures

GDPval-AA v2 evaluates models on realistic professional deliverables across white-collar occupations — analysis memos, financial models, legal summaries, project plans — scored by blind comparison against expert-produced work. Its Elo format captures head-to-head preference rather than pass/fail correctness, making it the closest published proxy for "would a professional accept this output?"

The Inversion

The notable part is the ordering: the $5/M-input model beats the $10/M flagship on exactly the workload category enterprises buy AI for. Anthropic's launch materials lean into the result, positioning Opus 5 for coding, research, and agentic knowledge work while reserving Fable 5 for long-horizon autonomous agents.

Enterprise buyers have noticed the same arithmetic that applies to coding workloads: for document-heavy professional work that completes within a session, the cheaper model is now also the published leader. Combined with Zapier AutomationBench results showing roughly 1.5× the pass rate of competitors on workflow automation, the enterprise pitch writes itself.

The Caveats

Elo leaderboards measure preference, not ground truth; blind comparisons can reward fluency and formatting alongside substance. And a 114-point gap over Fable 5, while meaningful, does not make the models categorically different — on long-horizon tasks the flagship still holds the recommendation.

The Bottom Line

Knowledge work is where AI spending is actually concentrating in 2026, and Opus 5 just took the published lead at half the flagship price. Expect the GDPval chart to appear in a lot of procurement decks this quarter.