Claude Haiku 5.5: Full Pricing and Benchmarks
Last updated: October 2026
Anthropic launched Claude Haiku 5.5 yesterday, October 7, and the headline number deserves the attention it is getting: the new small model is roughly 75% cheaper on average than Haiku 4.5, while more than doubling its knowledge-work score — 1620 vs 735 on GDPval-AA v2.1 — and jumping from 15.7% to 72.4% on the OSWorld 2.1 computer-use set.
Claude Haiku 5.5 (model ID claude-haiku-5-5) is Anthropic's cheapest, fastest, and most capable small model to date, priced at $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100k tokens. It targets high-volume, cost-sensitive work: summarization, classification, database queries, and context compaction.
Source: Claude official account on X
What Haiku 5.5 is, and who it is for
Anthropic's own definition is blunt: "the cheapest, fastest, and most capable small model we have released to date." The target workload is anything high-volume and cost-sensitive — summarization, context compactions, database queries, and classification — plus a role as a subagent inside coding work alongside Opus 5.5 and Sonnet 5.5, and fast direct customer support with browser use.
It is also the first Haiku-class model with adjustable effort settings — Low, Medium, High, Xhigh, and Max — so a quick classification and a harder subtask can run at different effort levels within the same pipeline. The Python and TypeScript SDKs ship with experimental support for computer-use and browser-use workflows.
Source: Anthropic
The official pricing, in full
Anthropic published a two-tier price list, with a sharp break at 100k tokens per request. All figures below are per million tokens, straight from the launch page.
| Item | Haiku 5.5 (up to 100k) | Haiku 5.5 (above 100k) | 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 reads | $0.01 | $0.05 | $0.10 | $0.10 (halved) |
| Cache writes | $0.125 | $0.625 | $1.25 | $2.50 |
In real terms, Anthropic says Haiku 5.5 is about 75% cheaper on average than Haiku 4.5: roughly 90% cheaper for requests under 100k tokens and 50% cheaper above that line. Since about 90% of Haiku 4.5 requests came in under 100k tokens, most existing traffic lands in the cheaper band.
One honest footnote from Anthropic itself: the new tokenizer consumes slightly more tokens per task, so effective savings depend on your workload's shape. The list price drop is dramatic; your invoice drop will usually be somewhat smaller than the sticker suggests.
Two more pricing moves shipped alongside the launch. Sonnet 5.5 cache reads were cut to $0.10, which Anthropic says makes agentic workloads about 20% cheaper. And Anthropic introduced new monthly API credits for subscribers: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team accounts.
The official benchmarks
The launch page publishes head-to-head numbers against Haiku 4.5, GPT-6 Luna, and Sonnet 5.5 as the reference point. This is the full table.
| Benchmark | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 (reference) |
|---|---|---|---|---|
| GDPval-AA v2.1 (knowledge work) | 1620 | 735 | 1437 | 1840 |
| AA-Briefcase v1.1 | 1578 | 614 | 1336 | 1824 |
| OSWorld 2.1 (computer use, offline set) | 72.4% | 15.7% | 48.9% | 83.9% |
| Humanity's Last Exam (no tools) | 45.9% | 10.2% | — | 56.9% |
| Humanity's Last Exam (with tools) | 57.4% | 18.7% | — | 64.5% |
| Terminal-Bench 4.0 (agentic coding) | 39.2% | 0.0% | 16.4% | 70.6% |
| FrontierCode 1.1 (Main) | 46.4% | — | 42.4% | 52.1% |
| Chartography | 46.4% | 6.4% | 29.1% | 61.6% |
Read the table two ways. Against its own predecessor, the jump is generational: more than double on GDPval-AA v2.1, roughly 4.5x on Humanity's Last Exam without tools (45.9% vs 10.2%), and from zero to 39.2% on Terminal-Bench 4.0 agentic coding.
Against OpenAI's small model, Haiku 5.5 outscores GPT-6 Luna on every benchmark where both appear — 1620 vs 1437 on GDPval-AA v2.1, 72.4% vs 48.9% on OSWorld 2.1, and 46.4% vs 42.4% on FrontierCode 1.1. But Sonnet 5.5 still sits clearly ahead of both, which is exactly the segmentation Anthropic describes.
Source: Anthropic
Speed you can measure, with named customers
Anthropic calls Haiku 5.5 its fastest model to date at standard speed — the Opus models in Fast Mode remain faster still. The more convincing evidence comes from named customers whose numbers are published on the launch page.
Asana cut response latency by more than 30% and measured inference speeds of up to 2.5x per turn after moving to Haiku 5.5. Box reports roughly half the response latency of Haiku 4.5 on its workloads.
For product teams, those are the numbers that matter more than benchmark points. Latency per turn compounds across every conversation, and a two-times improvement changes which features become viable to build at all.
Haiku 5.5, Sonnet 5.5, or Opus 5.5?
The decision rule Anthropic gives is worth repeating: Haiku 5.5 is for narrowly scoped, high-volume tasks, while Sonnet 5.5 and Opus 5.5 remain the best choice for complex agentic coding. Terminal-Bench 4.0 makes the gap visible — 39.2% for Haiku 5.5 against 70.6% for Sonnet 5.5.
A sensible split for a real product: Haiku 5.5 handles classification, extraction, summarization, and context compaction at scale, then hands hard reasoning and multi-file coding to Sonnet 5.5 or Opus 5.5. We cover both in our Claude Sonnet 5.5 and Claude Opus 5.5 launch reports.
What this means for you
For developers and startups worldwide — whether you are building in Lagos, Jakarta, or Manila — a capable model at $0.10 per million input tokens changes the unit economics of high-volume features that were previously too expensive to run. Support triage, document tagging, and search summarization all move from "maybe later" to "shippable now."
Haiku 5.5 is available at launch on AWS, Google Cloud, Microsoft Azure, and the Claude platform, so most existing infrastructure keeps working. If you also want Max-level access with the new API credits, a shared Claude Max 5x subscription runs €50 a month through the Truescho shop.
The competitive context matters here too. OpenAI spent the same week pushing GPT-6 to every ChatGPT tier — our GPT-6 Intelligent UI report covers it — while Anthropic answered on price and efficiency for developers, a contrast also visible across our OpenAI DevDay 2026 recap.
Quick comparison: Haiku 5.5 vs Haiku 4.5 vs Sonnet 5.5
| Haiku 5.5 | Haiku 4.5 | Sonnet 5.5 | |
|---|---|---|---|
| Input price (per 1M tokens) | $0.10 | $1.00 | $2.00 |
| Output price (per 1M tokens) | $0.50 | $5.00 | $10.00 |
| GDPval-AA v2.1 | 1620 | 735 | 1840 |
| OSWorld 2.1 (computer use) | 72.4% | 15.7% | 83.9% |
| Terminal-Bench 4.0 (agentic coding) | 39.2% | 0.0% | 70.6% |
| Adjustable effort | Yes — Low to Max | No | — |
| Best fit | High-volume, narrowly scoped tasks | Legacy small-model workloads | Complex agentic coding |
Source: Anthropic
The fine print
- The new tokenizer uses slightly more tokens per task, so real-world savings will be somewhat smaller than the raw price cut suggests.
- It is not the model for complex agentic coding — Anthropic itself points to Sonnet 5.5 and Opus 5.5 for that, with Haiku 5.5 cast as the subagent and high-volume workhorse.
- Cyber safeguards are stricter than Haiku 4.5 but looser than Anthropic's most recent models: defensive security tasks are allowed, penetration testing is not.
- Biological safeguards sit at the level of Sonnet 5 and 5.5 and Opus 5.
Community reaction on Hacker News — 893 points and 430 comments — leaned philosophical. Commenters debated an "LLM Moore's law" of roughly 90% efficiency gains every 18 months for the same quality over the past five years, and noted that small models tend to "forget" unhelpful information and re-derive it on the spot, trading memory for reasoning tokens.
Frequently asked questions
How much does Claude Haiku 5.5 cost per million tokens?
Requests up to 100k tokens cost $0.10 per million input tokens and $0.50 per million output tokens. Above 100k tokens, pricing rises to $0.50 input and $2.50 output. Cache reads start at $0.01 and cache writes at $0.125 per million tokens under 100k.
Haiku 5.5 vs Sonnet 5.5 — which should you pick?
Pick Haiku 5.5 for high-volume, well-scoped work such as summarization, classification, database queries, and context compaction. Pick Sonnet 5.5 for complex agentic coding and harder reasoning — it scores 1840 on GDPval-AA v2.1 against Haiku 5.5's 1620, at twenty times the input price.
Is Haiku 5.5 faster than GPT-6 Luna?
Anthropic publishes no head-to-head speed test against GPT-6 Luna. On Anthropic's own benchmarks, Haiku 5.5 outscores GPT-6 Luna on every test where both appear — 1620 vs 1437 on GDPval-AA v2.1 and 72.4% vs 48.9% on OSWorld 2.1 — and Anthropic calls it its fastest model at standard speed.
What is adjustable effort in Haiku 5.5?
Haiku 5.5 is the first Haiku-class model with tunable effort settings: Low, Medium, High, Xhigh, and Max. You choose how much effort the model puts into a given request, so a cheap quick task and a harder subtask can run at different levels inside the same workflow.
Is Haiku 5.5 available on AWS, Azure and Google Cloud?
Yes. At launch, Claude Haiku 5.5 is available on Amazon Web Services, Google Cloud, and Microsoft Azure, as well as on Anthropic's own Claude platform. The Python and TypeScript SDKs also add experimental support for computer-use and browser-use workflows at launch.
Our verdict
If you run any high-volume language workload on Haiku 4.5 today, move it to Haiku 5.5 this week — a roughly 75% average cost cut with a large capability gain is as close to a free lunch as this industry offers. Keep Sonnet 5.5 or Opus 5.5 for complex agentic coding, use Haiku 5.5 as the tireless subagent, and check the 100k token threshold against your real request sizes before modeling your bill. And if you want Claude's top tiers without the top-tier invoice, shared Claude Max subscriptions from the Truescho shop keep them affordable.
Sources
- Anthropic — Claude Haiku 5.5 — the official launch page with pricing, benchmarks, and customer results
- Anthropic — Claude Haiku — the official Claude Haiku model family page
- Anthropic — Claude — the full Claude model family overview
- Claude official account on X — the official Claude Haiku 5.5 announcement post