Playbooks · Coding tasks
Haiku 5.5 is now the default Haiku
Written November 9, 2026
Haiku 5.5 is the default Haiku — 1M context, $0.10/$0.50 per million tokens. What changes for existing Haiku users, and what doesn't.
What changed
Claude Code 2.1.293, released October 7, 2026, carries one line that changes what the word haiku means: "Added Claude Haiku 5.5 (claude-haiku-5-5), now the default Haiku model on the Anthropic API — 1M context, $0.10/$0.50 per Mtok ($0.50/$2.50 for prompts over 100K)." Anyone calling the haiku alias on the Anthropic API now gets Haiku 5.5, with no config change and no announcement inside the session. Anthropic's own Haiku page describes 5.5 as the fastest and most efficient model in the Claude 5.5 family, built for high-volume, cost-sensitive work, and the first Haiku with effort controls, so teams can tune cost against intelligence for each task.
The switch is a genuine upgrade on the numbers. Haiku 5.5 reads up to 1 million tokens at once, five times Haiku 4.5's 200,000, and writes up to 128,000 tokens in a reply. Anthropic positions it for the jobs a cheap model should do: classification, extraction, routing, summarization, compaction, and subagent work handed off by a larger model. It is available on Claude.ai on every plan, on the Claude Platform, in Claude Code, and through Amazon Web Services, Google Cloud, and Microsoft Foundry, under the model ID claude-haiku-5-5.
The price has a cliff at 100,000 tokens
Haiku 5.5 is the only current Claude model priced by prompt length, and the step is steep. For prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens. Past that line, the whole request reprices at $0.50 input and $2.50 output per million tokens — not just the tokens above the line, the entire request, including any part of the prompt served from cache. Anthropic's pricing docs state the rule plainly: a request whose prompt is over 100,000 tokens pays the higher prices, and each request is priced on its own, so an earlier short request keeps its lower price. Cache reads cost $0.01 per million tokens under the line and $0.05 above it, and the Batch API takes 50% off input and output. The headline rate is a tenth of Haiku 4.5's $1/$5, but a long prompt costs five times the headline rate, so a subagent loop that grows past 100K tokens stops being the cheap option mid-session.
What to do about it
- Find every place you wrote
haiku. Search settings files, subagent definitions, and scripts for the bare alias. Each one now resolves to a different model than it did before October 7, with a different context window, different output habits, and a different rate card. - Pin the full model ID where the choice matters. Write
claude-haiku-5-5where you want the new model and its 1M window, andclaude-haiku-4-5where you want the model you already tested. The changelog scopes the default change to the Anthropic API, so do not assume the alias moved on every provider you use — check your own setup before assuming. - Keep Haiku prompts short on purpose. Route narrow, verifiable work to Haiku — one lookup, one summary, one small fix with tests — and keep the prompt under 100,000 tokens. A subagent that accumulates tool output across dozens of turns will cross the line quietly; compacting the conversation or splitting the work across fresh subagents keeps the request on the cheap tier.
- Budget for the higher tier where prompts are long by design. If a job genuinely needs a long prompt on Haiku, price it at $0.50/$2.50 per million tokens from the start. The 1M window is real capacity, but it is priced for occasional use, not as a cheap 1M-token default.
Verdict
Pin the full model ID wherever the model choice matters, and keep Haiku prompts under 100,000 tokens wherever the price matters. The alias moved on one surface and the price jumps fivefold past 100K, so a bare haiku in a config file is now a decision you did not make. For short, scoped subagent work under the line, Haiku 5.5 is the right default; for long-context work, budget the higher tier or use a model whose full window is flat-rate.
What can go wrong
- Assuming the default moved everywhere. The changelog says "on the Anthropic API." A setup routed through another provider may still resolve
haikuto Haiku 4.5. The fix is one line: pin the full ID. - Letting a subagent loop cross 100K tokens unnoticed. The price step applies to the whole request, cache hits included, so a session that grows gradually gets five times dearer without any single call looking expensive.
- Treating the 1M window as free headroom. The window holds a million tokens, but every token past 100,000 in a request is billed at the higher tier. Capacity and price are different numbers on this model.
- Expecting Haiku to do Sonnet's job. Anthropic pitches Haiku 5.5 for narrow, repeatable work and pairs it with Opus or Sonnet as the planner. A hard coding task routed to Haiku because it is cheap is a mismatch, not a saving.
Limits
- This page reads Anthropic's Haiku page, its pricing docs, and the Claude Code 2.1.293 release notes. No API call was made here, so it reports no measured cost or latency of its own; the prices are Anthropic's list prices in US dollars.
- The default change is stated for the Anthropic API in the release notes. This page does not state what the
haikualias resolves to on any other provider, because the release notes do not say. - Prices change. The figures here are the ones on Anthropic's pricing page at the time of writing; the live pricing page governs.
Related playbooks
- Paperclip: native Runner now default
- Antigravity CLI 1.5.7: Auto becomes the default
- Claude Code 2.1.321: Fable 5.2 is default
- Haiku 5.5 is 90% cheaper only under 100K tokens: price the tier jump and the tokenizer before you route background work to it
Sources
- Claude Haiku — Anthropic
- Pricing — Claude Platform Docs
- Release v2.1.293 — anthropics/claude-code — GitHub
Back to: Coding playbooks hub · Also: Model prices: what the major models cost · Plan allowances: what each agent plan lets you do · Claude agents now dream: memory that consolidates