Skip to content

Model releases

Haiku 5.5 cuts token prices 90% on prompts under 100k

7 min read

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

All about: Claude prices and limits →
Video: x.com

Anthropic released Claude Haiku 5.5 on October 7 with a price list that splits at 100,000 tokens. Below that line, input costs $0.10 per million tokens, output costs $0.50 and cache reads cost one cent. Haiku 4.5 charged $1.00, $5.00 and $0.10 for the same three, at any prompt length.

Anthropic’s announcement calls the new model “the cheapest, fastest, and most capable small model we’ve ever released” and puts the average saving over Haiku 4.5 at around 75%. It is the first new Haiku in nearly a year, The News Stack noted. Haiku 4.5 shipped on October 15, 2025. The same outlet counts Haiku 5.5 as Anthropic’s third 5.5 model in a month and says Fable 5.5 will likely go through a much longer review process.

Two tiers and a new tokenizer

Above 100,000 tokens, Haiku 5.5 costs $0.50 per million input tokens and $2.50 per million output tokens. Cache reads in that tier cost $0.05, and cache writes cost $0.625, against $0.125 on shorter prompts and $1.25 on Haiku 4.5. The News Stack works the headline rates out to a 90% cut on short prompts and a 50% cut on long ones, and says the 75% average reflects a mix of requests plus an updated tokenizer that uses slightly more tokens per task. A tokenizer splits text into the units the model bills by, so the same prompt can now come to a slightly higher count.

Anthropic says prompts up to 100,000 tokens made up around 90% of requests to Haiku 4.5. That is where it says the new model is the best value. Even in the upper tier, Haiku 5.5 costs a quarter of Sonnet 5.5’s $2.00 input and $10.00 output. On its Haiku product page, Anthropic adds that prompt caching can save up to 90% and batch processing 50%.

Haiku 5.5 is also the first Haiku-class model with an adjustable effort setting. As on Anthropic’s larger models, it lets you trade intelligence against cost per task. The News Stack reports that the default is medium. The announcement charts accuracy against cost per attempt at each effort level on three benchmarks, including the offline subset of OSWorld 2.1, which tests agents operating a real computer through long multi-step tasks.

The benchmark table

Anthropic compares Haiku 5.5 with Haiku 4.5, OpenAI’s GPT-6 Luna and, for reference, Sonnet 5.5. The scores are Anthropic’s own runs, and the methodology is in the Haiku 5.5 system card. Figures are listed as Haiku 5.5 / Haiku 4.5 / GPT-6 Luna / Sonnet 5.5:

  • GDPval-AA v2.1 (knowledge work): 1,620 / 735 / 1,437 / 1,840
  • AA-Briefcase v1.1 (knowledge work): 1,578 / 614 / 1,336 / 1,824
  • OSWorld 2.1, offline subset (computer use): 72.4% / 15.7% / 48.9% / 83.9%
  • Humanity’s Last Exam, no tools: 45.9% / 10.2% / not reported / 56.9%
  • Humanity’s Last Exam, with tools: 57.4% / 18.7% / not reported / 64.5%
  • Terminal-Bench 4.0 (agentic coding): 39.2% / 0.0% / 16.4% / 70.6%
  • FrontierCode 1.1 Main (agentic coding): 46.4% / not reported / 42.4% / 52.1% for Sonnet 5.5 at Xhigh effort
  • Chartography, no tools (visual reasoning): 46.4% / 6.4% / 29.1% / 61.6%

Haiku 5.5 beats GPT-6 Luna on every row where both have a score and trails Sonnet 5.5 on all of them. The gap to Sonnet is widest on Terminal-Bench 4.0, where it is more than 31 points.

The News Stack also brought in models Anthropic left out of the table. Artificial Analysis currently scores Z.ai’s GLM-5.3-Flash at 1,647 on GDPval-AA v2.1 and 1,454 on AA-Briefcase v1.1, the outlet reports, against Anthropic’s own figures of 1,620 and 1,578 for Haiku 5.5. On price, it points to Alibaba’s Qwen3.7 Flash on its international service: $0.03 per million input tokens and $0.13 per million output for inputs up to 32,000 tokens, rising to $0.10 and $0.40 for inputs from 32,000 to 256,000 tokens.

What Anthropic wants it doing

Anthropic’s stated use cases start with high-volume, repetitive work: summaries, database queries, classification requests and compaction, which squeezes a long agent context into a shorter summary so the session can keep going. Because it is also Anthropic’s fastest model, the company points to latency-sensitive jobs like live customer support and browser use. For coding, it pitches Haiku 5.5 as a subagent alongside Opus 5.5 and Sonnet 5.5. The Haiku page describes the setup: a larger model such as Fable or Opus plans the work and hands subtasks to Haiku, which makes it practical to run many agents in parallel.

Anthropic also says where Haiku 5.5 should not go. Sonnet 5.5 and Opus 5.5 “remain better choices for complex agentic coding tasks” like those in Terminal-Bench 4.0. The company says Haiku 5.5 is best suited to more narrowly scoped tasks “that might otherwise have been cost-prohibitive with previous versions of Claude.”

The News Stack points out that summarization, classification and routing have always been the core jobs for small models, and that decision models like Jev are already doing well there at an even lower price. In the outlet’s reading, that may be why Anthropic emphasizes compaction, database queries and speed-sensitive agent work.

Asana is the one customer named in the announcement. Aaron Vinh, a staff software engineer there, said the company ran Haiku 5.5 through the eval suite for AI Teammates, its agent product, on tasks like triaging bugs, setting up projects and searching large portfolios for high-risk or overdue work:

“Compared with the model we use today, we saw over a 30% reduction in latency for task completions and up to 2.5x faster inference per agent turn. It’s a noticeably snappier experience.”

Cyber and biology limits

Anthropic says Haiku 5.5 improves on Haiku 4.5 on almost all of its alignment evaluations, with far fewer instances of misaligned behavior and a lower willingness to cooperate with misuse. Its cybersecurity safeguards are stricter than Haiku 4.5’s but somewhat looser than those on other recent models. They allow a wider range of defensive tasks than Sonnet 5.5’s safeguards and still block penetration testing and other techniques more likely to be used by attackers. The biology safeguards match those of Sonnet 5, Sonnet 5.5 and Opus 5: research biology questions are allowed, and requests Anthropic judges likely to cause harm are restricted. Organizations that need broader access can apply to the Life Sciences Verification Program or the Cyber Verification Program.

Availability, credits and SDKs

Haiku 5.5 is live on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure. On the Claude Platform the model ID is claude-haiku-5-5, and Anthropic has published a migration guide. Its Haiku page also lists the model in Claude.ai on every plan and in Claude Code. Anthropic is also adding beta support for computer use and browser use to its Python and TypeScript SDKs, and it names Haiku 5.5 as a good fit for both.

This week, Max and Team subscribers start getting monthly API credits for the Claude Platform. Max 5x users get $100 a month, Max 20x users get $200, and Team subscribers get up to $500 pooled across their users. The credits work with any model, and Anthropic says they are meant for building tools, apps and agents that call the API.

Cheaper cache reads on Sonnet 5.5

The other price change is on Sonnet 5.5, which launched on September 28 at Sonnet 5’s prices. Cache reads drop by half, from $0.20 to $0.10 per million tokens. That matches what Haiku 4.5 charged for cache reads. Anthropic says cache reads make up a large share of token consumption, so the cut brings Sonnet 5.5’s cost on most agentic tasks down by around 20%. Its Terminal-Bench 4.0 chart shows Sonnet 5.5’s cost per attempt moving left at the same score.

The cut began rolling out across platforms on launch day, The News Stack reports, though some existing Azure and Google Cloud customers will have to wait a few more days.

No comments yet. Be the first.

Join the conversation

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