Claude Haiku 5.5 ranks 2nd in intelligence but real costs spiral

After the release of Claude Haiku 5.5, Artificial Analysis published its evaluation: the family has 5 variants, with the strongest scoring 43 on the Intelligence Index, ranking 2nd among 179 models, at an input price of $0.10 per million tokens (output $0.50 for prompts under 100K). However, its value-for-money rank is just 43rd, at roughly $0.21 per task. On AA-Briefcase, Haiku 5.5 (max) reached 1578 Elo, with confidence intervals overlapping GPT-6 Astra (max). Multiple independent tests show its headline pricing is heavily eroded by abnormal token consumption and tiered pricing.

Confirmed

Why it matters

Haiku has always been positioned as the value option, but 5.5 combines first-tier intelligence with runaway token costs: 2nd in intelligence yet 43rd in value, with per-task costs several times peers. Add the 5x price cliff beyond 100K context and agentic tasks routinely exceeding 100k tokens, and hidden thinking/output costs can dwarf headline prices. Developers billed by token must benchmark real costs on their own workloads rather than trusting the price sheet.

2026-10-08 ~ 2026-10-09 · 16 related posts

Full story(3 episodes)→

Primary sources

1 near-duplicate retellings: RexDouglass