Claude Haiku 5.5 Arrives With a 90% Price Cut — Anthropic’s Cheapest, Fastest Small Model Yet

A glowing, stylised illustration of a compact AI chip and code stream representing Anthropic's Claude Haiku 5.5 small language model

Anthropic has launched Claude Haiku 5.5, its smallest and cheapest AI model, with pricing that undercuts its predecessor by up to 90%. Released on October 7, the model costs just $0.10 per million input tokens and $0.50 per million output tokens for most requests — a dramatic reduction from Haiku 4.5’s $1 and $5 rates. The launch also brings a one-million-token context window, an adjustable “effort” setting never before seen on the Haiku line, and benchmark scores that put it ahead of OpenAI’s rival small model, GPT-6 Luna.

But the headline number comes with fine print. Here’s what the price cut actually means, what got better besides the price, and why it matters for developers and startups in Pakistan.

Key Takeaways

  • Claude Haiku 5.5 costs $0.10/$0.50 per million tokens for requests under 100,000 tokens — a 90% cut from Haiku 4.5, matching OpenAI’s GPT-6 Luna.
  • The average saving is closer to 75%, because a new tokenizer consumes about 30% more tokens and longer prompts are priced 5x higher.
  • It ships with a 1M-token context window, an adjustable effort dial, and big benchmark jumps — 72.4% on OSWorld 2.1 vs 15.7% for Haiku 4.5.

The Price Cut, Explained

Haiku 5.5 is the only model in Anthropic’s lineup with two price tiers, split at a 100,000-token boundary:

Model Input / Output per 1M tokens Cache read Context window
Haiku 4.5 $1.00 / $5.00 $0.10 200K
Haiku 5.5 (up to 100K tokens) $0.10 / $0.50 $0.01 1M
Haiku 5.5 (over 100K tokens) $0.50 / $2.50 $0.05 1M
Sonnet 5.5 $2.00 / $10.00 $0.10 1M
Opus 5.5 $4.00 / $20.00 $0.20 1M

For requests under 100,000 tokens, the 90% reduction is real: input costs fall to one-tenth of Haiku 4.5’s rate. Anthropic says roughly 90% of Haiku 4.5 requests fell into this lower-cost category, so most users will see the full discount. Cross the 100,000-token line, however, and the rate jumps fivefold to $0.50/$2.50 — still half the price of Haiku 4.5, but no longer the 90% headline.

There is a second caveat. Haiku 5.5 uses a new tokenizer that produces roughly 30% more tokens for the same text, which trims the effective savings. Anthropic itself states the average saving is around 75%, not 90%. The cheaper token is not automatically a cheaper task.

What Got Better Besides the Price

This is not just a discount on an old model. Haiku 5.5 is a genuine generational jump for the Haiku line:

  • Bigger memory: The context window grew fivefold, from 200,000 to one million tokens, with maximum output doubled to 128,000 tokens.
  • Adjustable effort: For the first time in the Haiku series, users can set an effort level (low, medium, high, xhigh, or max — medium is the default) to trade cost against intelligence on a per-request basis.
  • Sharper cybersecurity posture: Anthropic says safeguards are stricter than Haiku 4.5’s, though lighter than Sonnet 5.5’s, with defensive security tasks permitted at larger scale.

The Benchmarks: Big Jumps, With a Caveat

Anthropic’s own test results show Haiku 5.5 leaping past its predecessor — and, on several measures, past OpenAI’s cheapest model, GPT-6 Luna:

  • OSWorld 2.1 (a test of operating real computer software): Haiku 5.5 scored 72.4%, versus 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna.
  • Terminal-Bench 4.0 (agentic coding): 39.2%, up from effectively zero on the prior Haiku.
  • FrontierCode v1.1 (real-world coding): 46.4%, ahead of GPT-6 Luna’s 42.4%.
  • GDPval-AA v2.1 (knowledge work): 1,620 points versus Haiku 4.5’s 735.
  • Humanity’s Last Exam: 45.9% without tools and 57.4% with tools.

The important caveat: all of these figures are vendor-reported. Independent benchmarks were still pending at launch. One independent analysis noted that Haiku 5.5’s maximum-effort mode generates roughly three times as many output tokens per task as GPT-6 Luna, which narrows the real-cost advantage on reasoning-heavy workloads.

Why It Matters in Pakistan

For Pakistani software houses, freelancers, and startups building AI products, the launch changes the cost math significantly. At current exchange rates (roughly Rs 277 to the dollar), one million input tokens — around 750,000 English words — now cost about Rs 28 on Haiku 5.5, down from roughly Rs 277. A customer-support chatbot handling thousands of queries a day, an AI tool summarising Urdu or English documents, or an agent extracting fields from invoices could see its API bill fall by three-quarters.

That matters because per-request cost has been the main reason many local projects stayed as demos rather than becoming products. When each AI answer costs a fraction of a rupee, high-volume use cases — support automation, content moderation, data extraction, classroom tools — finally pencil out.

The competitive angle matters too. Haiku 5.5’s price now exactly matches GPT-6 Luna’s ($0.10/$0.50), and Anthropic positioned the launch as a direct alternative to OpenAI’s cheapest tier. For Pakistani developers choosing a small model for an app, the two are now priced identically — the decision comes down to quality and features, not cost.

Two more practical notes:

  • The 100K cliff matters for big documents. Developers working with long Urdu or English texts should watch prompt sizes: one oversized request is repriced fivefold, so splitting documents into smaller chunks keeps you in the cheap tier.
  • Access is immediate. The model is live on the Claude Platform and through Amazon Web Services, Google Cloud, and Microsoft Azure, so Pakistani teams already using those clouds can switch without new infrastructure.

Anthropic also cut Sonnet 5.5 cache-read pricing in half to $0.10 per million tokens (roughly 20% cheaper agentic work, per the company) and began rolling out monthly API credits to paid subscribers — $100 for Max 5x users, $200 for Max 20x, and up to $500 pooled for Team plans.

What to Watch

The launch fits a clear industry pattern: AI answers are getting cheaper across the board as model choices multiply. OpenAI, Google, and Anthropic are all racing to own the high-volume, low-cost tier of the market, and that race is good news for anyone building at scale.

Still, cheaper tokens multiply whatever process they touch. A well-designed workflow gets genuine leverage from a 75% cost cut; a sloppy one just produces mistakes more cheaply. And the real test will be independent benchmarks over the coming weeks, which will confirm whether Haiku 5.5’s vendor-reported gains hold up outside Anthropic’s own lab.

For now, the headline stands: the fastest Claude got dramatically cheaper, noticeably smarter, and directly price-matched with its biggest rival. For Pakistan’s growing AI developer community, this is one of the most practical model launches of the year.