Benchmark Shift｜CursorBench: Claude Fable 5.1 Debuts at #1, Bumps Grok 4.6 to Third

TL;DR CursorBench 3.2: Fable 5.1 Max hits 73.4% (previous leader Grok 4.6 Extra High was 70.8%) and debuts by taking both first and second place; Fable 5.1 Max beats the old leader by 2.6 points while costing only $9.64 per task, 44% cheaper than the prior Fable 5 Max at $17.32; Anthropic's own site still lists only Fable 5, with no official Fable 5.1 announcement

#ai-agent #benchmark #daily #cursorbench #coding-agent

Table of Contents

Summary
Ranking Changes
1. CursorBench 3.2 — 2026-09-02
Analysis: What This Reshuffle Means
Today's Insight
References

Series: AI Benchmark Watch (1 / 1)

🌏 中文版

Summary

CursorBench — Cursor's own agentic-coding evaluation board — just had its biggest reshuffle since version 3.2: two entries labeled "Fable 5.1," Max (73.4%) and Extra High (72.8%), parachuted straight into first and second place, bumping the previous snapshot's (2026-08-31) leader Grok 4.6 Extra High (70.8%) down to third. Notably, Anthropic's own site and model documentation still list only Fable 5 as of this writing — no official Fable 5.1 announcement exists yet. This leaderboard entry arrived ahead of the official announcement.

Ranking Changes

CursorBench 3.2 — 2026-09-02

Rank	Model/Config	Score	Previous	Change
🥇	Fable 5.1 Max	73.4%	New entry	🆕
🥈	Fable 5.1 Extra High	72.8%	New entry	🆕
🥉	Grok 4.6 Extra High	70.8%	70.8% (🥇)	↓2
4	Fable 5 Max	70.5%	70.5% (🥈)	↓2
5	Opus 5 Max	70.0%	70.0% (🥉)	↓2

Source: cursor.com/cursorbench · Snapshot date: 2026-09-02

The scores for the former third-through-fifth place entries didn't change at all — they were simply pushed down by two new entries landing above them, not re-benchmarked.

Analysis: What This Reshuffle Means

Technical

The interesting part isn't how much Fable 5.1 won by — it's what it cost to win. CursorBench also publishes average cost per task: Fable 5.1 Max runs $9.64 per task, 44% cheaper than the prior Fable 5 Max's $17.32, using 30% fewer tokens (72,060 vs. 103,525) while still scoring 2.9 points higher (73.4% vs. 70.5%). Fable 5.1 Extra High is even more striking: $6.96 per task — over 60% cheaper than the old leader Fable 5 Max — while still outscoring it (72.8% vs. 70.5%). A same-generation point release (5 → 5.1) usually just nudges the score; getting both a score increase and a cost cut at the same time looks more like a training or routing change than simply throwing more compute at the problem.

Methodology

⚠️ One caveat is worth flagging clearly: as of this writing, neither Anthropic's own site (anthropic.com/claude/fable) nor its model docs (platform.claude.com) list anything beyond claude-fable-5 — there's no model card, pricing page, or official release announcement for Fable 5.1. The name comes entirely from community observation so far: a new model identifier spotted on Amazon Bedrock, early-access testing under the internal codenames "Melon" and "Marshmallow," and several tracker articles placing the release window between August 31 and early September. CursorBench is a third-party evaluation built and run by Cursor rather than the model vendor, which makes it more credible than a vendor self-report — but it's still currently the only public leaderboard showing this new version's score, with no second independent source cross-verifying it yet, and it can't be ruled out that this reflects an unfinalized early build.

Industry

If "Fable 5.1" is accurate, it means Anthropic is signaling a new model through a third-party leaderboard before making any official announcement — the same pattern seen before Opus 5's official launch, when it first leaked under the EAP codename "Honeycomb." For teams currently running coding agents on Fable 5 Max, the practical takeaway from this shift may not be "should we switch models" but rather "the next point release of the same family might deliver a higher score at a lower cost." If the eventual official pricing holds, Fable 5.1 Extra High would match or beat Fable 5 Max at less than half the cost — which would directly change the cost/benefit calculus for model selection.

Today's Insight

I used to assume a point release (5 → 5.1) mostly means a small score bump with pricing left unchanged. This CursorBench data suggests otherwise: a minor version update within the same model generation can cut cost and token usage while raising the score at the same time — beating the previous generation's most expensive configuration with a cheaper one. When judging whether a new version is worth switching to, checking the score alone isn't enough; cost per task and token usage per task matter too, or you'll underestimate how big a point release actually is.