Model results · Anthropic

Claude Haiku 5.5

Every published VulcanBench measurement of Claude Haiku 5.5, newest first: combined score, Code quality, runtime, tokens and Claude Code’s list-price cost with its price-tier floor and ceiling at every effort level on Frontier v4, graded by deterministic hidden tests on real tasks. Measured through Claude Code on a Claude Max subscription, with Claude Code’s refusal fallback on (no run used it). For the earlier Claude Haiku 4.5, see its own page; Claude Sonnet 5.5 and Claude Opus 5.5, the other Claude 5.5 models on the board, have their own pages too.

Suitev4

Claude Haiku 5.5 across every effort level

In Claude Code across five effort levels: combined score 84.42 at Low, 84.03 at Medium, 88.01 at High, 91.83 at Extra-high and 92.16 at Max (judged n=23, 22, 23, 22, 21). Tasks passed 13, 13, 17, 22 and 22 of 23; $0.84 to $1.54 a task as Claude Code reports it.

October 12, 2026 · 23 tasks · 5 effort levels · 115 runs of Claude Haiku 5.5 · VulcanBench Frontier v4 · Code quality protocol v3.24

→

Suitev4

Claude Haiku 5.5 vs Claude Sonnet 5.5, GPT-6.1 Sol and GPT-6 Luna

The four models side by side at every effort level on the same 23 tasks and judges, with three comparison cards. At Max Haiku 5.5 and Sonnet 5.5 are effectively tied (92.16 ± 0.37 against 92.02 ± 0.35), Haiku about 3.6 times cheaper per task.

October 12, 2026 · 23 tasks · 5 effort levels · 460 runs · VulcanBench Frontier v4

→

Suitev4

Cost, tokens and Haiku’s price tiers

Haiku 5.5 costs five times as much for prompts over 100,000 tokens, so each level shows Claude Code’s own total beside a floor and a ceiling. $1.04, $1.36, $1.25, $0.84 and $1.54 per task from Low to Max; the whole sweep priced at $138.31 (floor $31.10, ceiling $155.50).

October 12, 2026 · API-equivalent estimates, solver inference only · VulcanBench Frontier v4

→

How would Claude Haiku 5.5 do on your stack?

Public results answer "which model is stronger on this suite." Whether Claude Haiku 5.5 belongs in your routing table depends on your languages, your codebases, and your task mix, that gets measured, not guessed. Grading rules are in the methodology.

Optimize your model routing All reports