SemiAnalysis: Fable 5.2 tops its GDPx benchmark at 16.4% but costs 31% more per task
SemiAnalysis reports Fable 5.2 leads its GDPx professional-task benchmark at 16.4%, with cost per task rising from $67.16 to $88.15.
SemiAnalysis ranks Claude Fable 5.2 first on its GDPx professional-task benchmark at 16.4%, but the same analysis found that each task cost 31% more to run than on the previous model, rising from $67.16 to $88.15. The cost increase came despite a 20% cut in per-token prices, which fell from $10 to $8 per million input tokens and from $50 to $40 per million output tokens. The newer model costs less per token but uses more tokens per task.
The reason, according to SemiAnalysis's run logs as summarized by Latent Space, is how much checking the model does. Where Opus 5.5 and Fable 5.1 tried to write about 20 test cases, Fable 5.2 sometimes wrote hundreds or thousands. It also spawned subagents that re-derived results independently. SemiAnalysis credits that testing breadth for the benchmark lead. Latent Space's summary puts the cost point bluntly: you pay for the checks either way.
SemiAnalysis, through the account of its analyst @dylan522p, sorted the gains by domain. Fable 5.2 is significantly better at chip design, software engineering, biology and materials science, and not better at finance, legal work, quant trading, RSI or hardware design. The analyst said the RSI result reflects Anthropic's explicit nerf, meaning a deliberate limitation on that category of work rather than a capability gap. The domain split matters because a single headline benchmark number averages across tasks where the model improved and tasks where it did not.
The cost picture also depends on which effort setting is used. Artificial Analysis, as reported in the same roundup, found that at high and xhigh effort Fable 5.2 scores 62 using about 98k output tokens per task, against 173k for Opus 5.5 at max. At max, though, its cost per task is about $26.5, roughly 2.6 times Opus 5.5's $10.0 and twice Fable 5.1's $12.99. Anthropic's own FrontierCode 2.0 claim is that Fable at High matched Opus 5.5 at Max for about half the cost per task. Several evaluations show max and xhigh underperforming high, and @Yuchenj_UW advised against using Fable 5.2 Max.
The GDPx result is one leaderboard entry among several. On Artificial Analysis's Intelligence Index, Fable 5.2 scores 65 at max effort, ahead of Gemini 4 Carbon at 60 and Opus 5.5 at 59. Its weaker spots include AA-Briefcase, where it sits within the confidence interval of Opus 5.5, and GDP.pdf, where it posted 24% all-pass against 26% for Opus 5.5. The mixed results fit the cost picture. The model does more work per task, and that extra work does not translate into gains everywhere.
What changed
Fable 5.2 per-token prices fell from $10 to $8 per million input tokens and from $50 to $40 per million output tokens, while SemiAnalysis's GDPx cost per task rose from $67.16 to $88.15.
Type: Pricing and limits · Applies to: Fable 5.2
Why it matters
Per-token price cuts do not guarantee lower project costs when a model spends more tokens on verification. Buyers should budget using task-level costs from their own workloads.
What you can do
Track cost per completed task, not per-token price, when deciding which model to deploy for coding or research agents.
Original source
We have no link to the original source for this story yet.
Where we saw it
Latent Space
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works