Claude Haiku 5.5 cuts prices 90% to $0.10 per million tokens and beats the previous Sonnet on office work

What happened

On 7 October Anthropic released Claude Haiku 5.5, the third and smallest model in its Claude 5.5 family, after Opus 5.5 (22 September) and Sonnet 5.5 (28 September). Its predecessor in the small tier is Haiku 4.5.

Anthropic calls it the cheapest, fastest and most capable small model it has released. It is aimed at high-volume, cost-sensitive work: summaries, compressing long conversations, database queries and sorting requests. It can also act as a helper inside tasks led by a larger model, and suits jobs that need quick responses such as live customer support and browser use. It is the first Haiku with an adjustable effort setting.

API price, prompts ≤100k tokens (per million in / out)

$0.10 / $0.50

Per-token price vs Haiku 4.5

−90%

Average running cost vs Haiku 4.5

about −75%

Context window

1M tokens

Sources: Anthropic, “Introducing Claude Haiku 5.5”; Claude API pricing and models overview pages

How the price works

Haiku 5.5 has two price tiers by prompt length: if a request’s input, including cached input, exceeds 100,000 tokens, the whole request is billed at the higher rate.

Per million tokens	Haiku 5.5 (≤100k / >100k)	Haiku 4.5	Sonnet 5.5	Opus 5.5
Input	$0.10 / $0.50	$1.00	$2.00	$4.00
Output	$0.50 / $2.50	$5.00	$10.00	$20.00
Cache reads	$0.01 / $0.05	$0.10	$0.10	$0.20

Sources: Anthropic’s announcement and the Claude API pricing page, checked 9 October 2026.

Anthropic says about 90% of requests to Haiku 4.5 were under 100,000 tokens, so most users will pay the lower rate. Above the threshold the saving over Haiku 4.5 is 50%. The “about 75% cheaper on average” figure also accounts for Haiku 5.5’s new tokenizer, which uses slightly more tokens for the same work.

Using the published prices, here are a few rough costs (our estimates, ignoring thinking tokens and caching discounts; real bills vary with the effort setting):

• Sorting 10,000 emails into categories, at about 800 tokens per email including instructions and a 50-token answer: about $1.05 on Haiku 5.5, $10.50 on Haiku 4.5 and $21 on Sonnet 5.5.
• Summarising a 300,000-token document (above the threshold) into 2,000 tokens: about $0.16 on Haiku 5.5 and $0.62 on Sonnet 5.5.

How much better is it?

Anthropic compares it with Haiku 4.5, OpenAI’s small GPT-6 Luna and its own Sonnet 5.5:

• Claude Sonnet 5.51840
• Claude Haiku 5.51620
• GPT-6 Luna1437
• Claude Haiku 4.5735

Each model at its highest effort setting. Source: Anthropic, “Introducing Claude Haiku 5.5”, 7 Oct 2026

Test	What it measures	Haiku 5.5	Haiku 4.5	GPT-6 Luna	Sonnet 5.5
AA-Briefcase v1.1	Long-running knowledge work (Elo)	1578	614	1336	1824
OSWorld 2.1 (offline subset)	Using a computer like a person	72.4%	15.7%	48.9%	83.9%
Humanity’s Last Exam (no tools)	Expert-written questions	45.9%	10.2%	—	56.9%
Terminal-Bench 4.0	Completing jobs in a command line	39.2%	0.0%	16.4%	70.6%
FrontierCode 1.1	Whether code changes could be merged	46.4%	—	42.4%	52.1%
Chartography	Reading charts (no tools)	46.4%	6.4%	29.1%	61.6%

Source: Anthropic’s announcement. “—” means not reported.

Some yardsticks:

• It beats the previous mid-tier model. On GDPval-AA, Haiku 5.5’s 1620 is above Sonnet 5’s 1449; on FrontierCode its 46.4% beats Sonnet 5’s 42.4% (Sonnet 5 figures from the Sonnet 5.5 announcement). Per token, Haiku 5.5 costs a twentieth of Sonnet 5.
• Work Haiku 4.5 could not do at all. Haiku 4.5 scored 0% on Terminal-Bench 4.0, failing every task; Haiku 5.5 scores 39.2%.
• Still well behind Sonnet 5.5. On command-line coding it trails by 31 points, 39.2% to 70.6%.

What a task actually costs

Anthropic published each model’s score against its real cost per task at each effort level. This is what decides whether a small model is worth using:

Test	Model (effort)	Score	Cost per task
GDPval-AA	Haiku 4.5 (Max, its best)	735	$0.24
	Haiku 5.5 (Low)	1125	$0.012
	Haiku 5.5 (High)	1420	$0.089
	GPT-6 Luna (Max, its best)	1437	$0.09
	Sonnet 5.5 (High)	1551	$0.62
OSWorld 2.1	GPT-6 Luna (Medium)	37.5%	$0.13
	Haiku 5.5 (Medium)	53.3%	$0.13
	Haiku 5.5 (Max)	72.4%	$0.61
	Sonnet 5.5 (Low)	57.9%	$0.68
Terminal-Bench 4.0	Haiku 5.5 (Max)	39.2%	$2.64
	Sonnet 5.5 (High)	43.0%	$1.46

Source: cost charts in Anthropic’s announcement. Sonnet 5.5’s Terminal-Bench cost uses the cache price in force from 7 October.

• Against Haiku 4.5: at its lowest setting, Haiku 5.5 scores nearly 400 points more than Haiku 4.5’s best, for about a twentieth of the cost.
• Against OpenAI: for the same $0.13 per computer-use task, Haiku 5.5 scores about 16 points more than GPT-6 Luna; on office work the two are level at about $0.09.
• Smaller is not always cheaper. On Terminal-Bench, Haiku 5.5 at maximum effort scores 39.2% for $2.64; Sonnet 5.5 at High scores 43% for $1.46. Anthropic says plainly that Sonnet 5.5 and Opus 5.5 remain the better choice for complex coding.

What early testers report

From customers quoted in Anthropic’s announcement; not independently checked:

• AlphaSense runs about 8 million calls a week through its document Q&A feature, one of its biggest costs. On 400 queries Haiku 5.5 scored 0.84 against Haiku 4.5’s 0.76, a statistically significant gain.
• Box saw scores 11 points higher than Haiku 4.5 at about half the latency.
• Asana measured more than 30% lower latency for completed tasks and up to 2.5 times faster inference per agent turn.
• Cognition says that with Opus 5.5 leading and Haiku 5.5 as its helper, its Devin tool scores 66.2 on FrontierCode while cutting cost and latency.

Two other changes announced the same day

• Sonnet 5.5 cache reads halved, from $0.20 to $0.10 per million tokens. Anthropic says cache reads are a large share of token use in agent tasks, so most of them get about 20% cheaper.
• API credits for Max and Team subscribers. Max 5x gets $100 a month and Max 20x $200; Team gets $20 per Standard seat and $100 per Premium seat, pooled and capped at $500. At Haiku 5.5’s lower rate, $100 buys about a billion input tokens, or nearly a million emails sorted by the estimate above. According to the Help Center, you claim the credit by linking a Claude Console organisation in claude.ai’s billing settings. It expires each cycle, does not cover interactive Claude Code, and does not raise chat or Cowork limits. Free, Pro and Enterprise plans are not eligible.

Things to keep in mind

• The vendor reports both scores and costs. GPT-6 Luna’s results were run or quoted by Anthropic, possibly under different conditions. Haiku 5.5 came out after our latest LMArena snapshot, so there is no blind-vote ranking yet.
• The same figure differs slightly between announcements. Sonnet 5.5’s GDPval-AA score is 1844 in its own announcement and 1840 in Haiku 5.5’s.
• Long prompts cost five times as much. Requests over 100,000 tokens are billed at $0.50 / $2.50, so check before processing large documents in bulk.
• Tighter cybersecurity limits than Haiku 4.5. It allows more defensive tasks than Sonnet 5.5 but blocks penetration testing and other techniques more likely to be used by attackers.

How to use it

• Developers. The model ID is claude-haiku-5-5, available on the Claude Platform and on Amazon’s, Google’s and Microsoft’s clouds, with effort defaulting to Medium. Anthropic also added beta computer-use and browser-use support to its Python and TypeScript SDKs, and says Haiku 5.5 is especially well suited to these tasks.
• Everyday users. The announcement covers only API and cloud availability; Claude’s pricing page lists Haiku-tier models on every individual plan from Free to Max.
• What to give it. High-volume sorting, extraction and summaries, and look-ups inside projects led by Opus or Sonnet. Rogo gives a typical example: while a bigger model builds a slide deck, a Haiku 5.5 helper goes into the company’s annual report and pulls out the segment revenue figure the deck needs.