Fable 5.2 ranks first in Agent Arena after 4,700 agent sessions

• Fable 5.2 (Max) took first place after 4,700 sessions, with 30.4% net improvement in Arena's announcement, up from Fable 5.1's 14.8%.
• Its advantage over Opus was concentrated in praise and steerability; Arena's comparison showed only a 1.3-point edge in confirmed success, and lower Bash reliability.
• The early task-cost saving has reversed: petergostev's reply put Fable at $7.30 per task, while the Nov 11 leaderboard lists $7.87, slightly above Fable 5.1.

Arena's success rubric requires an explicit answer to a completion pop-up; “perfect, thanks” alone earns no confirmed-success credit. Its cost calculation samples at most the first three tasks in each session to limit accumulated-context costs.

Fable 5.2 (Max) posted an eye-catching generation jump: 14.8% → 30.4% net improvement, +15.6 points, moving from its predecessor's fifth-place position to first in the announcement snapshot.

It also led all three category leaderboards:

• Code: +34% net improvement, up from sixth place for Fable 5.1.
• Work: +31%, up from fourth.
• Chat: +27%, up from sixth.

Fable led four of the five signals in Arena's comparison. The per-signal net-improvement scores below read Opus 5.5 (Max) → Fable 5.2 (Max):

• Praise vs. complaint: 40.3% → 73.4%, +33.1 points.
• Steerability: 18.6% → 33.0%, +14.4 points.
• Confirmed success: 20.7% → 22.0%, +1.3 points.
• Bash reliability: 15.1% → 14.4%, −0.7 points.
• Tool hallucination: 0.7% → 1.0%, +0.3 points, with higher scores indicating better performance.

Praise and steerability account for roughly 87% of the summed five-signal lead in that snapshot.

The current signal definitions contain some consequential scoring details:

• Praise: Tasks with equal praise and complaints, including no explicit sentiment, are excluded from this signal.
• Steering: Every correction counts against the model; repeated corrections incur a larger penalty.
• Bash: Correctly executed tests that discover bugs count as successful calls. Errors outside the agent's control are exempt.
• Tools: Invented tool names and nonexistent signatures count as hallucinations; Arena says base rates are low across frontier models.

Median task costs shifted between the announcement snapshot and the Nov 11 leaderboard:

• Fable 5.2 (Max): $7.30 → $7.87.
• Fable 5.1 (Max): $7.70 → $7.74.
• Opus 5.5 (Max): $5.50 → $5.59.

Fable 5.2's initial saving over 5.1 had disappeared in the later sample. Token pricing remained 20% lower: Arena's announcement reported a reduction from $10/$50 to $8/$40 per million input/output tokens.

Arena's cost methodology segments sessions into individual tasks, which can span multiple turns. Its Pareto view supports p25, p50 and p95 task costs, alongside output-token efficiency.

Users see neither the model identity nor its cost, according to petergostev's reply. Under Arena's causal-tracing methodology, the evaluation:

1. Randomly assigns an orchestrator model to each session.
2. Estimates its effect against a fixed baseline distribution, typically uniform over component choices.
3. Averages the five per-signal effect estimates into the headline score.

The production leaderboard currently evaluates orchestrator choice only, with one randomized component, or K=1. It applies time-decaying weights to emphasize recent observations and reports 95% confidence intervals.

By Nov 11, Fable's sample had grown to 7,841 sessions. Its overall estimate was 30.48% with a 95% confidence margin of ±3.07 points, versus Opus at 19.51% ±2.04 points, in the updated leaderboard.