First Third-Party Scores for Claude Fable 5.2: Cursor Says It Tops CursorBench at 60% on Max Effort and 56.9% at Default (Fable 5.1: 49.2%); Perplexity Reports a Record 0.851 on WANDR at $14.51 per Task

Background and Pain Points

When Anthropic released Claude Fable 5.2 early on November 10 SGT, its FrontierCode 2.0 and other scores were all self-reported, with no third-party data yet. For developers, evals run by integrators in real product settings are often more useful: Cursor's CursorBench draws on real coding tasks inside the editor, and Perplexity's WANDR measures how well its Computer agent completes tasks. Within an hour of the launch, both companies shipped Fable 5.2 and posted results.

Architecture Highlights

Neither company disclosed model details beyond their evals; the model changes come from Anthropic: pricing drops to $8/$40 per million tokens, output is up to 30% faster, and Medium is the default effort inside Claude products. Cursor's chart labels High as Fable 5.2's default effort in Cursor. The CursorBench 4.0 chart plots average cost per task on the x-axis and links each model's effort levels into a curve, so you can see how much score each extra dollar buys; Perplexity's WANDR chart likewise plots score against cost per task and traces a Pareto frontier.

Benchmarks

CursorBench (Cursor's numbers): Fable 5.2 scores 56.9% at default High effort, versus 49.2% for Fable 5.1, and 60% at max effort, which Cursor calls the top score on the board. The chart shows the default at about $9.70 per task and the max-effort point near $33, so roughly three times the cost buys about three more points. Curves for Gemini 4 Argon, Opus 5.5, Sonnet 5.5, Haiku 5.5, Muse Wonder, Grok 4.7 and GPT-5.6 Sol all sit below Fable 5.2, but Cursor did not post their exact values. WANDR (Perplexity's numbers): Fable 5.2 scores a record 0.851 at $14.51 per task. Along the frontier are Gemini 4 Argon 0.802 ($7.88), GPT-6.1 Sol 0.654 ($3.70), Claude Opus 5.5 0.647 ($3.69), GPT-6 Sol 0.566 ($2.23) and DeepSeek V4 Pro 0813 0.359 ($0.75); off the frontier are GPT-6 Astra 0.735 ($13.34), Claude Fable 5.1 0.684 ($13.16), Claude Sonnet 5.5 0.550 ($2.73) and Grok 4.7 0.528 ($9.30).

Getting Started

Cursor users can pick Fable 5.2 in the model selector now; the CursorBench curve suggests the default effort captures most of the gain, with max effort best saved for a few hard tasks. In Perplexity, Fable 5.2 is available in Computer for Pro and Max subscribers and in Search for Max only. If cost matters, WANDR shows Gemini 4 Argon reaching 0.802 at about half the cost per task, and Opus 5.5 reaching 0.647 at about a quarter. Both evals are vendor-built, with task sets and scoring not fully public, and results depend on each product's integration, so re-test on your own workloads before switching.