Claude Sonnet 5.5 at max reasoning writes worse code than xhigh while burning most tokens

DeepLearningAI · x · 2026-10-22

DeepLearning.AI's review of Claude Sonnet 5.5 and Haiku 5.5:

Counterintuitive finding: Sonnet 5.5 produced less mergeable code at max reasoning than at xhigh, one level below, while consuming more tokens than any model measured — "max reasoning ≠ best results." Test your settings before trusting them.
Haiku 5.5: first update in a year, scores 43 on Artificial Analysis' Intelligence Index (13th), 237 tokens/sec, at a steep price increase.
Sonnet 5.5: #1 on the Coding Agent Index (68), #2 on the Intelligence Index (56) at launch, nearly matching Opus 5.5 at Sonnet 5 prices, positioned for routine coding and office work.
Five adaptive reasoning levels, 1M-token input, June 2026 knowledge cutoff.
The Terminal-Bench 4.0 chart shows accuracy rising with cost per attempt, with max-level gains flattening; Haiku 5.5 beats Haiku 4.5 across levels.

More from Models