Claude Haiku 5.5 costs a tenth of Haiku 4.5 per token on most prompts

Anthropic has released Claude Haiku 5.5, the small, fast model in its Claude 5.5 family, and priced it far below Haiku 4.5 [1][3]. For prompts up to 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens; Haiku 4.5 costs $1.00 and $5.00 [1]. Anthropic says that, on average, the new model costs around 75% less to run, once you allow for a new tokenizer that uses slightly more tokens for the same work [1]. On the same day it halved the price of cache reads on Sonnet 5.5 and announced a monthly API credit for Max and Team subscribers [1]. For anyone who runs summaries, classification or subagents through Claude at volume, this is a price change worth re-doing the sums for.

What was announced

Anthropic calls Haiku 5.5 “the cheapest, fastest, and most capable small model we’ve ever released” [1]. It is built for high-volume, cost-sensitive work: summaries, compactions, database queries and classification requests, the company says [1]. Anthropic also pitches it as a subagent that works alongside Opus 5.5 and Sonnet 5.5 on coding work, and, because it is the company’s fastest model at standard speed, for speed-sensitive tasks such as live customer support and browser use [1]. A footnote adds that it runs less quickly than Opus models in Fast Mode [1].

The price has two tiers. Up to 100,000 tokens of prompt, Haiku 5.5 costs $0.10 per million input tokens, $0.50 per million output tokens, $0.01 per million for cache reads and $0.125 per million for cache writes [1]. Above 100,000 tokens, those figures become $0.50, $2.50, $0.05 and $0.625 [1]. Anthropic says the model is priced 90% lower than Haiku 4.5 for requests up to 100,000 tokens and 50% lower above it, and that about 90% of requests to Haiku 4.5 fell into the shorter category [1]. The average “around 75% less” figure accounts for the tokenizer change [1].

That tokenizer change is the detail to watch. Anthropic says Haiku 5.5 uses an updated tokenizer, similar to Sonnet 5.5’s and Opus 5.5’s, which “uses slightly more tokens per task” [1]. Simon Willison ran the same long prompt through his token counter and found it used around 1.25x as many tokens on Haiku 5.5 as on Haiku 4.5, which he calls “a hidden price increase” [2]. It does not cancel the cut, but it means the real saving on your work will be smaller than the headline per-token ratio.

The new model is also the first Haiku with an adjustable effort setting, so users can trade cost against intelligence, the company says [1]. Willison reports that Haiku 5.5 does not let you disable reasoning and defaults to medium effort [2]. In his own test, a low-effort run of his standard drawing prompt cost 0.0936 cents and took 7 seconds, while a max-effort run took 5 minutes 9 seconds and cost 3.3826 cents [2]. Latent Space, citing Artificial Analysis, reports that at max effort the model uses about 162k output tokens per Intelligence Index task, roughly 3x GPT-6 Luna’s ~50k [4]. Effort level, in other words, can move your bill more than the token price does.

Anthropic published a benchmark table comparing Haiku 5.5 with Haiku 4.5, OpenAI’s GPT-6 Luna and Sonnet 5.5 [1]. On OSWorld 2.1, a computer-use test, Haiku 5.5 scores 72.4%, against 15.7% for Haiku 4.5, 48.9% for GPT-6 Luna and 83.9% for Sonnet 5.5 [1]. On Terminal-Bench 4.0, an agentic coding test, it scores 39.2%, against 16.4% for GPT-6 Luna and 70.6% for Sonnet 5.5 [1]. On Humanity’s Last Exam without tools it scores 45.9%, against 10.2% for Haiku 4.5 [1]. Latent Space reports that Artificial Analysis gives Haiku 5.5 a score of 43 on its Intelligence Index at max effort, up 26 points on the previous Haiku and ahead of GPT-6 Luna at 38 [4]. These are the company’s own results, plus one third-party index; none of them is a test of your work.

The launch page carries customer quotes, chosen by Anthropic. HubSpot says Haiku 5.5 scored 92.8%, averaged over three runs, on its simulated CRM test suite, the best score it has seen there [1]. AlphaSense says that on 400 queries for its “Ask in Document” feature, Haiku 5.5 scored 0.84 against Haiku 4.5’s 0.76 [1]. Box says the model scored 11 points higher than Haiku 4.5 at about half the latency [1]. Asana reports over a 30% reduction in latency for task completions compared with the model it uses today [1].

Availability is broad. Haiku 5.5 is available on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure, and developers can call it on the Claude Platform as claude-haiku-5-5 [1]. AWS says it is available on Amazon Bedrock and on Claude Platform on AWS, and that Bedrock keeps data within AWS infrastructure with regional data residency [3]. Latent Space reports it is also live in Claude Code [4]. Anthropic is also updating its Python and TypeScript SDKs to add computer use and browser use in beta, and says Haiku 5.5 is well suited to those tasks [1].

What changed, and for whom

The first group affected is anyone paying for Claude by the token on short, repetitive work. If your prompts stay under 100,000 tokens, the per-token price drops by 90% [1]. Willison notes that this exactly matches the $0.10/$0.50 price of OpenAI’s GPT-6 Luna, released last month, which Haiku 4.5 cost a full 10x more than [2]. Above 100,000 tokens the comparison flips: Haiku 5.5’s price rises 5x, while Luna’s rises only to $0.20/$0.75 beyond 272,000 tokens, so for long prompts Willison says Luna “looks like a much better deal” [2]. Prompt length is now the first thing to check before choosing between them.

The second group is Sonnet 5.5 users. From the launch day, Sonnet 5.5’s cache reads cost 50% less: $0.10 per million tokens rather than $0.20 [1]. Because cache reads make up a large share of token use, Anthropic says this makes Sonnet 5.5 around 20% cheaper on most agentic tasks [1]. Its input and output prices stay at $2.00 and $10.00 per million tokens [1]. Anthropic is explicit that Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding, and that Haiku 5.5 suits narrowly scoped work such as compaction, summarisation and subagent tasks [1].

The third group is Claude subscribers. This week Anthropic is rolling out a monthly API credit for Max and Team plans: $100 a month for Max 5x, $200 for Max 20x, and up to $500 for Team subscribers, pooled across their users [1]. The credits can be used on any model [1]. Willison says claiming them means choosing, under Settings and Billing, which API organization should receive them; that the credits exactly match the cost of the subscription; and that they do not roll over [2]. He also points out that the API lets you turn off auto-reload so requests stop when the balance runs out, which caps the risk of a surprise bill [2]. Latent Space reports that the credits also work in third-party harnesses [4].

The last change is on safety. Haiku 5.5’s cybersecurity safeguards are more restrictive than Haiku 4.5’s but somewhat less restrictive than those on other recent Claude models; they allow a wider range of defensive tasks than Sonnet 5.5’s safeguards but still block penetration testing, the company says [1]. Its biology safeguards are the same as those on Sonnet 5, Sonnet 5.5 and Opus 5 [1]. Organisations doing wider-ranging work can apply to Anthropic’s Life Sciences and Cyber Verification Programs [1].

Who it is for — and not

Haiku 5.5 is for teams and solo builders who send many small, well-defined requests through Claude: tagging support tickets, summarising documents, pulling a figure out of a report, cleaning records, or running as a subagent under a bigger model. That is the work Anthropic designed it for, and the customer quotes on the launch page all describe tasks of that shape [1]. AWS describes the same split: a larger model plans and makes judgment calls, while Haiku 5.5 handles fast, high-volume subtasks such as routing, classification and small code changes [3].

It is also for Claude Max and Team subscribers who have wanted to try the API without opening a second bill. The new monthly credit gives them room to build small tools and agents, as long as they spend it within the month [1][2].

It is not a replacement for Sonnet 5.5 or Opus 5.5 on hard, open-ended work. On Terminal-Bench 4.0, Anthropic’s own table puts Haiku 5.5 at 39.2% against Sonnet 5.5’s 70.6%, and the company says the larger models remain the better choice for complex agentic coding [1].

It is not automatically the cheapest option for long prompts. Above 100,000 tokens Haiku 5.5’s price rises 5x, and Willison judges GPT-6 Luna the better deal there [2]. And it is not cheap at every setting: a max-effort answer can use several times the output tokens of a low-effort one [2][4]. Pick the lowest effort level that gives acceptable results.

Finally, it is not a reason to skip testing. Benchmarks and customer quotes are chosen by the vendor, and the tokenizer change means your token counts will move [1][2]. The only number that matters is cost per acceptable answer on your own work.

This page will be updated if Anthropic changes Haiku 5.5’s pricing or safeguards, when the subscriber API credits finish rolling out, and if independent tests confirm or contradict the benchmark and cost claims.

∴ Small-model work on Claude is about 90% cheaper per token under 100,000 tokens; re-price your high-volume jobs.