Cresta-Bench: leaderboard

A model runs four contact-center agents taking calls from an LLM-simulated caller across 181 cases, each run three times at the model's lowest reasoning setting, except Claude Sonnet 5.5 at low; a case counts only when all three runs pass, from Cresta's own runs. It rewards getting the same case right every time, not once.

Metric: Cases passed on all 3 runs, mean over 4 agents (%). Source: fireworks.ai. Saturation forecast: Around February 2027. 14 models tracked. Updated 16 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 5.5Low82.57#5
2Claude Sonnet 5.5Low78.04#19
3GPT-6 SolNon-reasoning77.94#26
4Gemini 3.8 FlashLow76.14#36
5GPT-5.4Non-reasoning65.14#128
6Gemini 3.6 FlashMinimal50.19#75
7GPT-5.2Non-reasoning49.47#178
8Qwen 3.8 MaxNon-reasoning48.67#59
9GLM-5.3Low43.06#68
10GPT-5.6 LunaNon-reasoning38.81#122
11Kimi K3Low34.85#56
12GPT-6 LunaNon-reasoning32.73#99
13DeepSeek V4.1 Flash (1017)Non-reasoning32.64#41
14Claude Haiku 5.5Non-reasoning20.05#70

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #6 GPT-6 Astra.

Interactive version: theaggregate.ai/benchmark?slug=cresta-bench · How It Works · Data refreshed daily, snapshot 2026-11-16.