The direct answer: Claude Haiku 5.5 is Anthropic’s new small AI model, launched on 7 October 2026 at $0.10 per million input tokens and $0.50 per million output tokens — 90% below Haiku 4.5, and the same rate card as OpenAI’s GPT-6 Luna (per Anthropic’s announcement). On Anthropic’s own tests it scores 72.4% on the OSWorld computer-use benchmark and tops the Artificial Analysis small-model index. Our verdict: buy it for high-volume routine work, but treat it as a routed workhorse — set the effort dial yourself and keep Sonnet 5.5 on standby for anything that needs real judgment.
What Claude Haiku 5.5 is, in plain terms
Anthropic’s Claude family has three sizes: Opus is the heavyweight, Sonnet is the everyday workhorse, and Haiku is the cheap, fast sprinter. Haiku 5.5 is the newest member of the line, and Anthropic calls it the cheapest, fastest and most capable small model it has ever released. It reads text and images, produces text, takes a 1-million-token context window, writes up to 128,000 tokens per answer, and is aimed at the jobs you run thousands of times a day: summarisation, classification, extraction, database queries, support triage, and subagent work under a bigger model.
It is available in the Claude apps (web, iOS, Android) on every plan from Free upward — free-tier users now get two current-generation models, since Sonnet 5.5 joined the free plan in late September, while Opus 5.5 stays paid-only (per Notebookcheck’s launch coverage). Developers reach it through the Claude Platform API as claude-haiku-5-5, and through AWS, Google Cloud and Microsoft Azure. One feature is new for a Haiku: an effort setting (low, medium, high, xhigh, max) that trades cost against intelligence — the first time the small model has had a thinking dial at all.
If you are choosing between a subscription seat and the metered API, read our Claude vs ChatGPT subscription value breakdown first — Haiku 5.5 sits on the API side of that decision, and it only makes sense once you know what the seats already give you.
The price card: a 90% cut with a step in it
Here is the pricing that matters, per Anthropic’s announcement (7 October 2026, rates per million tokens):
| Model | Input | Output | Cache read |
|---|---|---|---|
| Claude Haiku 5.5 (prompts up to 100K tokens) | $0.10 | $0.50 | $0.01 |
| Claude Haiku 5.5 (prompts over 100K tokens) | $0.50 | $2.50 | $0.05 |
| Claude Haiku 4.5 | $1.00 | $5.00 | $0.10 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $0.10 |
| Claude Opus 5.5 | $4.00 | $20.00 | $0.20 |
Two things in this table deserve more attention than the headline. First, the 100,000-token step: any single request whose prompt passes 100,000 tokens is billed at five times the rate on the whole request, not just the excess. Second, Anthropic’s own estimate is that Haiku 5.5 costs about 75% less to run than Haiku 4.5 on average — not 90%. The headline quotes token prices; the real number accounts for a newer tokenizer that counts roughly 30% more tokens for the same text, and thinking that is now on by default. Honest of Anthropic to publish the 75% figure next to the 90% one; most launches would not.
The Batch API halves these rates ($0.05/$0.25 for Haiku 5.5 under the line), and there is a second lever in the same announcement: Sonnet 5.5 cache reads were halved from $0.20 to $0.10, which Anthropic says makes Sonnet about 20% cheaper on most agentic work. Max 5x subscribers now get $100/month in API credits, Max 20x get $200, and Team plans up to $500 pooled — small teams can prototype routing setups on house money.
If reasoning tokens are still fuzzy for you, our reasoning-token pricing explainer walks through exactly what “thinking on by default” does to a bill.
What your tasks actually cost: the Haiku 5.5 calculator
Plug in your real numbers. The calculator prices your workload on Haiku 5.5, Sonnet 5.5 and Opus 5.5 side by side, applies the 100K-token step automatically, and honours a batch-discount toggle.
Formula: monthly cost = tasks × ((fresh input × input rate + cached input × cache rate + output × effort multiplier × output rate) ÷ 1,000,000) × (0.5 if batch). Haiku 5.5 uses the 5x rate card when the average prompt exceeds 100,000 tokens. Rates are Anthropic’s list prices (7 October 2026); the effort multipliers are illustrative examples, not Anthropic figures — this is our calculation.
Where the benchmarks actually put it
All of the following launch-day scores are Anthropic’s own numbers, and they chose to put GPT-6 Luna in the same table:
- OSWorld 2.1 (operating a computer): Haiku 5.5 72.4%, Haiku 4.5 15.7%, GPT-6 Luna 48.9%, Sonnet 5.5 83.9%
- Terminal-Bench 4.0 (command-line work): Haiku 5.5 39.2%, Haiku 4.5 0.0%, Luna 16.4%, Sonnet 5.5 70.6%
- GDPval-AA v2.1 (professional knowledge work): Haiku 5.5 1620, Haiku 4.5 735, Luna 1437, Sonnet 5.5 1840
- FrontierCode 1.1 (coding): Haiku 5.5 46.4%, Luna 42.4%, Sonnet 5.5 52.1%
- Humanity’s Last Exam (expert reasoning): Haiku 5.5 45.9% without tools / 57.4% with tools, versus 10.2% / 18.7% on Haiku 4.5
The jump over Haiku 4.5 is the story — Terminal-Bench going from 0.0% to 39.2% means an entire category of agentic work moved down a pricing tier. Independent testing broadly agrees on direction: Artificial Analysis scores Haiku 5.5 at 43 on its Intelligence Index at max effort, ahead of GPT-6 Luna at 38, making it the top small model in that index (per The GenAI Magazine’s review and Token Harbor’s comparison). Artificial Analysis also clocks it at 137 tokens/second at high effort and 243 at max — genuinely the fastest Claude.
Two honesty notes. First, vendor benchmarks flatter; Anthropic does not say whether it ran Luna itself or used OpenAI’s figures. Second, the price is per token and the tokens are not free: at max effort Haiku 5.5 writes roughly three times as many output tokens per task as Luna, so the per-task cost gap is smaller than the rate-card gap (Artificial Analysis via Token Harbor). Bito ran the three Claude 5.5 models through real coding sessions on launch day and found Haiku 5.5 at high effort scoring 79% at $0.62 per session, against Sonnet 5.5 at medium scoring 77% at $0.64 — matching the middle model, for less, on their workload (Bito’s independent test). That is one vendor’s workload, not a universal result, and on hard terminal tasks Sonnet’s 70.6% vs 39.2% is a chasm. Anthropic itself steers complex agentic coding to Sonnet and Opus.
And one published counter-data point worth keeping: Decrypt gave the model a quick logic question at launch and got a fast answer that was wrong — the price of speed on the wrong task.
Before trusting any index, run the model on your own documents — our guide to comparing AI models on your own work lays out the 20-task method.
The two gotchas nobody quotes in the headline
1. The tokenizer and the thinking eat the saving. Haiku 5.5 uses a newer tokenizer that counts about 30% more tokens for the same text than Haiku 4.5’s did, and thinking is on by default with medium as the default effort. That is the entire gap between the “90% cheaper” headline and Anthropic’s “75% cheaper on average” estimate. Dial the effort to low for classification and routing jobs and you get most of the headline back; leave it on max for everything and you will not.
2. The 100K step changes the whole request. Cross 100,000 tokens in a single prompt and every token in it bills at $0.50/$2.50. Agent loops that accumulate conversation history can drift across that line silently — the threshold counts the whole prompt, including cached input. If you run long-horizon agents, see our AI API spend limits guide for the guardrails that keep this from becoming a bill-shaped surprise.
Migration is not a drop-in swap either. Anthropic’s migration guide notes that thinking budgets, assistant prefill, custom temperatures and the old computer-use tool now return errors on Haiku 5.5, and new safety refusals can fire on requests Haiku 4.5 answered. Plan a short migration, not a model-name swap. The one durability fact in your favour: Anthropic says Haiku 5.5 will not be retired sooner than 7 October 2027, so you have a year of runway before a forced move.
Who should switch, and who should stay put
Switch if: you pay per token for classification, extraction, summarisation, support triage or routing at volume; you run subagents under Sonnet or Opus (Cognition already runs Haiku 5.5 as a “sidekick” to Opus 5.5 in Devin); you are on Haiku 4.5 and your prompts stay under 100K tokens — the savings are immediate. A no-code support setup like the ones in our AI customer support system guide is the classic fit: our calculation puts 20,000 tickets a month (1,500 input tokens, 300 output) at about $6.00 on Haiku 5.5 versus $60.00 on Haiku 4.5 and $120.00 on Sonnet 5.5.
Stay put if: one wrong answer is expensive — legal, financial or compliance judgment still belongs to Sonnet or Opus; your prompts routinely cross 100K tokens (at that point you are paying $0.50/$2.50 against Sonnet’s flat $2/$10, and the quality gap may no longer be worth it); your Haiku 4.5 integration hardcodes thinking budgets or prefill (budget migration work, then move); or you only use the Claude app casually — free-tier users already get Haiku 5.5 and Sonnet 5.5 at $0, so the API math is irrelevant until you start building.
The pattern the evidence supports is routing, not replacement: let Haiku handle the well-scoped volume and escalate ambiguous or high-impact cases to Sonnet or a human. Our AI agents for business piece covers where that routing pattern works — and where the agent idea is oversold.
Run the 20-task test on your own work
Do not migrate on benchmarks. Take 20 real tasks from your own pipeline — ten easy, five medium, five genuinely ambiguous — and run each through Haiku 5.5 at medium effort and Sonnet 5.5. Score correctness with the same rubric, record the token counts from the API response headers, and compute cost per correct task, not per token. If Haiku clears your bar on the easy and medium sets, route them there and keep Sonnet for the ambiguous five. That is the whole decision, and it takes an afternoon. For the API-vs-seat half of the question, our ChatGPT API vs Plus analysis shows the same metered-vs-flat arithmetic on the other side of the fence.
What we cannot confirm
We did not hands-on test this model. Every benchmark above is either Anthropic’s or a third party’s; the newest independent figures (Artificial Analysis, Bito) were collected days after launch, and Anthropic’s Luna comparison carries the vendor’s choice of methodology. OpenAI’s GPT-6 Luna list price is quoted via MarkTechPost as reported by secondary coverage — we read Anthropic’s page directly on 19 October 2026, and the price table there is the primary source for every Haiku and Sonnet figure in this article. The “30% more tokens” tokenizer figure comes from Anthropic’s technical documentation as reported by secondary trackers; if your token counts differ, the migration guide is the authoritative place to check.
Is Claude Haiku 5.5 free?
The API is pay-as-you-go with no free plan, but the model itself is selectable in the Claude apps on every plan including Free — you just cannot call the API for free.
How much cheaper is Claude Haiku 5.5 than Haiku 4.5?
Token prices are 90% lower ($0.10/$0.50 vs $1/$5 per million), but Anthropic’s own average estimate is about 75% cheaper to run, because the new tokenizer counts ~30% more tokens and thinking is on by default.
Is Claude Haiku 5.5 better than GPT-6 Luna?
Same rate card ($0.10/$0.50), but Anthropic’s table has Haiku 5.5 ahead of Luna on every shared benchmark, and Artificial Analysis independently scores it 43 vs 38. The catch: Haiku 5.5 writes roughly three times as many output tokens at max effort, so per-task costs are closer than per-token prices suggest.
Should I use Claude Haiku 5.5 or Sonnet 5.5?
Haiku 5.5 for high-volume, well-scoped work — classification, extraction, routing, subagents. Sonnet 5.5 for complex coding and anything where a wrong answer is expensive; it scores nearly double on Terminal-Bench 4.0 (70.6% vs 39.2%).
Does thinking stay on by default?
Yes — adaptive thinking is on by default with medium as the default effort, and it delays the first token at higher settings. Drop it to low for classification-style jobs to recover most of the headline saving.
What happens above 100,000 tokens in one prompt?
The whole request bills at 5x: $0.50 input / $2.50 output per million tokens. Anthropic says about 90% of Haiku 4.5 requests fell under the line; agent loops with growing history are the main way teams cross it.
Shurah is the founder of AI Tools Daily, tracking pricing, licensing and policy changes across AI tools so readers can make decisions without wading through marketing claims themselves.
