Skip to content

Anthropic

Short prompts to Claude Haiku 5.5 cost 90% less

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: AI prices →
Image: shattered.io

Claude Haiku 5.5 went on sale on October 7 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens, a tenth of the $1/$5 that Haiku 4.5 charges, per Anthropic’s Haiku 5.5 page. Classification, routing and summarization traffic made of short calls now pays 90% less per token, but the saving per task is smaller: The News Stack reports that Anthropic puts the average at around 75%, partly because an updated tokenizer uses slightly more tokens for the same task.

Prompts over 100K tokens move to a second tier of $0.50/$2.50. That is still half of Haiku 4.5’s rate, which was the same whatever the request size. Per The News Stack, Anthropic says about 90% of requests to Haiku 4.5 fell into the lower-priced band, and the 75% average accounts for that mix of requests as well as the tokenizer change. Prompt caching saves up to 90% on top of the list price, and batch processing saves 50%.

Bigger window, new effort setting

Anthropic’s model documentation lists a 1M-token context window and 128K maximum output for Haiku 5.5, against 200K and 64K for Haiku 4.5. Haiku 5.5 has a June 2026 knowledge cutoff.

The other cost lever is new for this tier. Haiku 5.5 is the first Haiku with effort controls, a setting for how many tokens the model spends on a task, and the default is medium. On Haiku 4.5, extended thinking worked through a manual budget_tokens value with no effort setting; Haiku 5.5 uses adaptive thinking instead. A pipeline ported from Haiku 4.5 therefore has a new setting to choose, or to leave at medium.

How it scores against Sonnet 5.5

The News Stack published a benchmark table that sets Haiku 5.5 against Haiku 4.5, OpenAI’s GPT-6 Luna and Sonnet 5.5:

  • Terminal-Bench 4.0 (agentic coding): Haiku 5.5 39.2%, Haiku 4.5 0.0%, GPT-6 Luna 16.4%, Sonnet 5.5 70.6%
  • OSWorld 2.1, offline subset (computer use): 72.4%, 15.7%, 48.9%, 83.9%
  • FrontierCode 1.1 Main: Haiku 5.5 46.4%, GPT-6 Luna 42.4%, Sonnet 5.5 52.1% at Xhigh effort; no Haiku 4.5 score
  • GDPval-AA v2.1 (knowledge work): 1620, 735, 1437, 1840
  • Humanity’s Last Exam without tools: Haiku 5.5 45.9%, Haiku 4.5 10.2%, Sonnet 5.5 56.9%; with tools 57.4%, 18.7% and 64.5%

Haiku 5.5 beats GPT-6 Luna on every benchmark in the table where both have a score. Sonnet 5.5 still scores almost twice as high on Terminal-Bench. On price the order flips: Sonnet 5.5 kept Sonnet 5’s $2/$10 rates in September, and Haiku 5.5’s short-prompt rates are a twentieth of that. Anthropic splits the work along that line. Opus 5.5 is its “daily driver” for complex coding and knowledge work, and Sonnet 5.5 fits well-scoped tasks. Haiku 5.5 is for subagents, high-volume text work and real-time chat, voice and live support. In Anthropic’s suggested setup, Fable or Opus plans the work and hands subtasks to Haiku so that many agents can run in parallel.

The News Stack adds that the classic small-model jobs of summarization, classification and routing are where decision models like Jev are currently making a splash, at an even lower price.

What early customers measured

Anthropic’s page carries results from customers’ own evaluations, each on that customer’s tasks:

  • HubSpot runs models through simulated CRM portals and mostly tests smaller models. Haiku 5.5 scored 92.8% averaged over three runs, which HubSpot calls the best result it has seen on that suite. HubSpot says it was also the fastest model on a stale-records audit task, with the highest hit rate and the lowest false-positive rate.
  • Box says Haiku 5.5 scored 11 points higher than Haiku 4.5 at about half the latency in early testing.
  • A team running an “Ask in Document” feature at about 8M calls a week tested 400 queries and got 0.84 for Haiku 5.5 against 0.76 for Haiku 4.5. The team calls the gap statistically significant.
  • A team behind an agent product called AI Teammates measured a latency cut of over 30% on task completion, compared with the model it uses now, and up to 2.5x faster inference per agent turn.
  • Devin Fusion, with Opus 5.5 as the lead and Haiku 5.5 as the “sidekick”, holds a FrontierCode score of 66.2, its developers say, while cutting cost and latency.

Getting it

Per The News Stack, Haiku 5.5 is Anthropic’s third 5.5 model in a month and the first new Haiku in nearly a year. Fable 5.5 is still missing from the family, and it will likely go through a much longer review process. Haiku 4.5 came out on October 15, 2025.

Haiku 5.5 runs on Claude.ai for Free, Pro, Max, Team and Enterprise users on web, iOS and Android, and in Claude Code. Developers can use it on the Claude Platform and through Amazon Web Services, Google Cloud and Microsoft Foundry; the tiered prices quoted above are Claude Platform rates. In the API the model is claude-haiku-5-5, and Anthropic’s documentation now lists Haiku 4.5 as Legacy, with advice to consider migrating to Haiku 5.5.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.