Ten cents. That is what one million input tokens cost on Claude Haiku 5.5, the model Anthropic released on October 7, 2026 — a 90% cut from Haiku 4.5’s $1.00. For anyone running AI at volume, this is the cheapest serious model Anthropic has ever shipped.
But the headline price has a trapdoor: Anthropic’s own fine print says the real-world saving is smaller. Here are the official numbers and the margin math.
What Anthropic actually announced
Anthropic calls Haiku 5.5 its cheapest, fastest, and most capable small model to date. It is built for high-volume, cost-sensitive work — summaries, compactions, database queries, classification — and pairs with Opus 5.5 and Sonnet 5.5 as a subagent on coding work. Its speed also suits live customer support and browser use.
The official spec sheet:
- 1M-token context window (up from 200K on Haiku 4.5) with 128K max output tokens
- Adjustable effort setting — low, medium, high, xhigh, max, defaulting to medium. The first Haiku model with an effort dial.
- Model ID
claude-haiku-5-5— a fixed ID with no date suffix - Knowledge cutoff of June 2026; available on the Claude API, AWS, Google Cloud, and Azure
Anthropic also halved Sonnet 5.5’s cache-read price and added monthly API credits: $100 (Max 5x), $200 (Max 20x), up to $500 pooled (Team) — usable on any model.
The official pricing, in one table
Haiku 5.5 has two price rows, split at 100,000 prompt tokens:
| Per 1M tokens | Haiku 5.5 (≤100K prompt) | Haiku 5.5 (>100K prompt) | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|---|
| Input | $0.10 | $0.50 | $1.00 | $2.00 |
| Output | $0.50 | $2.50 | $5.00 | $10.00 |
| Cache read | $0.01 | $0.05 | $0.10 | $0.10 |
| Cache write (5-min) | $0.125 | $0.625 | $1.25 | $2.50 |
The 100K-token cliff: the gotcha in the fine print
Here is the trapdoor: if your prompt exceeds 100,000 tokens, the entire request reprices at the higher tier — $0.50 input and $2.50 output per million.
Three details from Anthropic’s own docs make this cliff easier to hit than it looks: prompt length counts all input tokens, including cache reads and writes; the new tokenizer produces roughly 30% more tokens for the same text; and adaptive thinking is on by default, with thinking tokens billed.
That is why Anthropic’s own headline is not 90%. The company says Haiku 5.5 costs around 75% less to run on average than Haiku 4.5 — a vendor estimate that bakes in the cliff and the hungrier tokenizer. The 90% figure is real, but only for prompts under 100K tokens, which Anthropic says made up about 90% of requests to the previous Haiku model.
How the pricing cliff works
Prompt ≤ 100,000 tokens
$0.10 input / $0.50 output per 1M. The headline price. ~90% of Haiku workloads land here.
Prompt > 100,000 tokens
$0.50 input / $2.50 output per 1M — the whole request reprices 5x. Cache tokens count toward the limit.
How good is it? The benchmark evidence
Anthropic’s own benchmarks — vendor figures, not independent proof — show a large jump over Haiku 4.5:
- OSWorld 2.1 (computer use): 72.4% vs 15.7% for Haiku 4.5, and ahead of GPT-6 Luna’s 48.9%
- Terminal-Bench 4.0 (agentic coding): 39.2% vs 0.0% for Haiku 4.5 — though Sonnet 5.5 still leads at 70.6%
- FrontierCode 1.1: 46.4% vs 42.4% for GPT-6 Luna
Anthropic is explicit: Sonnet 5.5 and Opus 5.5 remain better for complex agentic coding. Haiku 5.5 targets narrowly scoped work that used to be cost-prohibitive.
How to profit: sell high-volume AI work at Haiku prices
This is where the 90% cut turns into a business: automation services that were marginal before — anything repetitive, high-volume, and narrowly scoped — are now profitable.
- Pick a repetitive, high-volume job. Inbox triage, lead qualification, review summarization, data extraction, support-desk first replies.
- Do the margin math with official rates. Lead qualification: 1,500 input tokens in, 200 out. Cost per lead: (1,500 ÷ 1,000,000 × $0.10) + (200 ÷ 1,000,000 × $0.50) = $0.00025. At 5,000 leads a day, that is about $37.50/month. Charge a client $299/month and you keep roughly $261 — an ~87% gross margin.
- Engineer every prompt under 100K tokens. Chunk long documents. Cache your system prompt — reads cost $0.01 per million. Never drift over the cliff.
- Stack the Batch API for non-urgent work. Overnight reports and bulk classification get another 50% off: $0.05 input / $0.25 output per million.
- Use the effort dial as a cost control. Run routing and classification at low effort; reserve high or max effort for steps where quality affects revenue.
- Prototype on subscriber credits first. Max 5x includes $100/month in API credits, Max 20x $200, Team plans up to $500 pooled — all usable on any model. Validate before spending real API budget.
Find work priced by human hours, deliver it at ten-cent-per-million-token cost, and keep the spread.
Keep Reading
FAQ
What is the Claude Haiku 5.5 model ID?
claude-haiku-5-5 — a fixed ID with no date suffix, working across the Claude API, Bedrock, Google Cloud, and Azure.
When was Claude Haiku 5.5 released?
October 7, 2026 — the third release in the Claude 5.5 family, after Opus 5.5 and Sonnet 5.5.
What happens if my prompt exceeds 100,000 tokens?
The whole request bills at the higher tier: $0.50 per million input, $2.50 per million output. Prompt length includes cache reads and writes.
How much cheaper is it really than Haiku 4.5?
90% cheaper per token under 100K ($0.10 vs $1.00 input). Anthropic estimates ~75% cheaper on average across real workloads — its own figure, accounting for the cliff and a hungrier tokenizer.
Does the Batch API discount apply?
Yes — 50% off input and output: $0.05/$0.25 per million under 100K tokens. Ideal for bulk, non-urgent jobs.
Is Haiku 5.5 good enough for coding?
For narrow, checkable tasks and subagent work, yes. For complex multi-step agentic coding, Anthropic still recommends Sonnet 5.5 or Opus 5.5.
Do Claude subscribers get API credits to try it?
Yes: $100/month on Max 5x, $200 on Max 20x, up to $500 pooled on Team plans — usable on any model.