Claude Haiku 5.5: The Fast Tier Grew Up

Haiku 4.5 shipped in October 2025 and spent a year as the oldest model in the Claude lineup, with the smallest window and no current coding scores we could verify. Haiku 5.5 replaces it on almost every axis at once:

Data verified October 9, 2026 against Anthropic's Haiku 5.5 model page, migration guide, launch post, and the Claude Code changelog.

This page never had verified coding scores to show for Haiku 4.5. Haiku 5.5 comes with them. These are the vendor-published numbers from Anthropic's launch post, next to the previous Haiku, the current Sonnet, and OpenAI's equivalent fast tier:

Two readings. Against its own lineage the jump is enormous: 0.0% to 39.2% on Terminal-Bench 4.0 is the difference between a model you cannot hand a terminal and one you can. Against Sonnet 5.5 the gap is still large, 31 points on Terminal-Bench, which is exactly why this stays the fast tier and not a cheaper default. All figures are Anthropic's own; we have not reproduced them.

Every multiple below is computed from list prices for prompts up to 100K tokens:

A worked example. A triage pipeline labels 10,000 issues a day at 2,000 input and 200 output tokens each, 20M input and 2M output tokens measured on Haiku 4.5:

Haiku 4.5   20M x $1 + 2M x $5     = $30.00 / day
Sonnet 5.5  20M x $2 + 2M x $10    = $60.00 / day
Haiku 5.5   20M x $0.10 + 2M x $0.50  = $3.00 / day
  + ~30% more tokens (new tokenizer)     = $3.90 / day

Even after the tokenizer uplift that is about 8x cheaper than the same pipeline on Haiku 4.5. Two honest caveats: thinking tokens bill as output, so run high-volume routes at low effort, and long prompts cost more. One 150K-token prompt with a 2K answer costs $0.08 on Haiku 5.5 versus $0.16 on Haiku 4.5, the roughly 50% saving Anthropic quotes for long requests rather than 90%.

Haiku 5.5 is a much better model than Haiku 4.5, and that makes the old failure mode more tempting, not less. Multi-file features, debugging where the cause sits far from the symptom, architecture calls and long unattended runs still need the reasoning depth of Sonnet or Opus 5.5. A 31-point Terminal-Bench gap to Sonnet 5.5 shows up as review time.

In Claude Code, /model haiku selects Haiku 5.5 from v2.1.293, and the picker shows the effort ladder. On older versions the alias still points at Haiku 4.5, so update the CLI first. On the API, it is a new model ID plus a few breaking changes:

{
  "model": "claude-haiku-5-5",
  "max_tokens": 16000,
  "thinking": { "type": "adaptive" },
  "output_config": { "effort": "low" },
  "messages": [{ "role": "user", "content": "Label this issue: ..." }]
}

Full details in Anthropic's Haiku 5.5 migration guide. Forced tool_choice still works, but the response then skips thinking; use auto when you want the model to reason before calling the tool.

The best Haiku deployment is still a heterogeneous fleet. A lead session on Claude Opus 5.5 owns the plan and the review and fans bounded subtasks out to Haiku 5.5: searches, per-file summaries, mechanical changes. Each subagent call costs 1/40 of the same output on Opus 5.5, so the fan-out can go wide while the judgment stays with the flagship. With a 1M window on both sides, the subagent no longer works from a truncated view of the repo.

The same logic applies one level up. On an AIDEN board the model is picked per feature card: the mechanical cards (rename sweeps, doc syncs, lint burn-downs) run on Haiku 5.5 through your own Claude Code install, while Sonnet and Opus take the cards that need a real engineer's model. For the full routing playbook, see model routing for coding agents.