Claude Haiku

Part of Anthropic on Models.

The fast, cheap Claude — for sub-agents, quick checks and high-volume steps.

Claude Haiku is the smallest of the Claude lineup, and a great little fighter: fast at standard speed, a tenth of Sonnet's input price, and — on Haiku 5.5 — a Terminal-Bench 4.0 score no previous Haiku came near. Anthropic's own positioning is high-volume, latency-sensitive work: classification, routing, extraction, subagent tasks.

What it is good at

• Sub-agents and quick checks — the work where waiting costs more than thinking
• High-volume steps at $0.10 in / $0.50 out per MTok for prompts up to 100k tokens
• Anthropic's fastest model to date at standard speed (vendor-reported)

What it is less good at

• Not the model for the hardest agentic coding — Terminal-Bench 4.0 39.2% against Sonnet 5.5's 70.6% (vendor-reported)
• Prompts over 100,000 tokens step up to $0.50 in / $2.50 out per MTok

Reaching it with TeamCLI

• Claude Code lists Claude Haiku on its own model screen, so a seat with model: Claude Haiku runs it the same way the Opus and Sonnet seats run theirs.
• Cursor and Antigravity list it among the Claude models on their plans.

Versions

One block per version, the current one first, each with its own anchor. The comparison table puts the costs and the family's benchmark columns side by side; the values are verbatim, and every one is sourced and dated in its version's section below and in the Sources list. A dash means the source publishes no figure, and nothing here is estimated.

No independent intelligence or speed index is republished here: Artificial Analysis measures both across labs, and its terms allow the link but not the copy — its figures live on its own pages.

Claude Haiku 5.5 — current

• Model id: claude-haiku-5-5

Abilities

• Released: October 7, 2026 — Anthropic — Claude Haiku 5.5 model page (read 2026-10-19).
• Context window: 1M tokens — Anthropic — Claude Haiku 5.5 model page (read 2026-10-19).
• Max output: 128K tokens — Anthropic — Claude Haiku 5.5 model page (read 2026-10-19).
• Output speed: —
• Latency: Fastest (comparative); Anthropic's fastest model to date at standard speed — Anthropic — Claude Haiku 5.5 announcement (read 2026-10-19).

Costs

• Input price: $0.10 / MTok — Anthropic — pricing (read 2026-10-19). Prompts over 100,000 tokens: $0.50 in / $2.50 out
• Output price: $0.50 / MTok — Anthropic — pricing (read 2026-10-19).

Benchmarks — as reported by the lab

• Terminal-Bench 4.0 (agentic coding) — 39.2% (Anthropic — Claude Haiku 5.5 announcement, read 2026-10-19)
• FrontierCode 1.1, Main (agentic coding) — 46.4% (Anthropic — Claude Haiku 5.5 announcement, read 2026-10-19)
• GDPval-AA v2.1 (knowledge work, Elo) — 1620 (Anthropic — Claude Haiku 5.5 announcement, read 2026-10-19)
• Humanity's Last Exam — 45.9% no tools / 57.4% with tools (Anthropic — Claude Haiku 5.5 announcement, read 2026-10-19)
• OSWorld 2.1 (computer use) — 72.4% (offline) (Anthropic — Claude Haiku 5.5 announcement, read 2026-10-19)

Claude Haiku 4.5

• Model id: —

Abilities

• Released: —
• Context window: —
• Max output: —
• Output speed: —
• Latency: —

Costs

• Input price: $1 / MTok — Anthropic — pricing (read 2026-10-19).
• Output price: $5 / MTok — Anthropic — pricing (read 2026-10-19).

Benchmarks — as reported by the lab

• Terminal-Bench 4.0 (agentic coding) — 0.0% (Anthropic — Claude Haiku 5.5 announcement, read 2026-10-19)

Sources

• Anthropic — Claude Haiku 5.5 model page — read 2026-10-19
• Anthropic — pricing — read 2026-10-19
• Anthropic — Claude Haiku 5.5 announcement — read 2026-10-19

Figures refreshed 2026-10-19; benchmarks are vendor-reported. Prices are per million tokens unless a figure says otherwise.