Claude Haiku 5.5: Token Pricing, API Credits, and What to Hand Off in Practice
Claude Haiku 5.5's official pricing around 100K tokens, Max and Team API credits, benchmarks, and subagent use, plus what the 75% cost cut means.
Joohwan Yu · Founder, Full Stack Developer
Toris · Published
Contents
Anthropic announced Claude Haiku 5.5 on October 7, 2026. A post introducing it on Arca Live prompted me to check the official announcement. The core of the announcement is running high-throughput, cost-sensitive work faster and cheaper.
Some tasks, like summarization, context compaction, and classification, have small individual requests but a lot of repetitions. For these tasks, the time to finish a request and the cumulative cost matter as much as the model’s peak performance. Anthropic introduces Haiku 5.5 as a model tuned for this area, and for coding work it emphasizes its role as a subagent that helps Opus 5.5 and Sonnet 5.5.
This post sums up the pricing and performance figures from the official announcement and looks at how to apply it to product development. The benchmarks below are evaluation results published by Anthropic, not measurements Toris (토리스) made by running the model ourselves. Prices and subscription benefits are based on the official announcement as checked on October 8, 2026.
The Work Haiku 5.5 Targets
The main uses named in the official introduction are as follows.
- Summarization and compaction of long documents or conversations: keeping the key context needed for later work.
- Database queries and classification requests: handling scoped questions repeatedly.
- Coding assistant agents: handling small subtasks within work owned by a larger model.
- Real-time customer support and browser use: interactions where a fast response matters.
The model ID is claude-haiku-5-5. The official announcement says it is available on platforms including the Claude Platform, AWS, Google Cloud, and Microsoft Azure. For the detailed changes when moving between environments, check the official migration guide.
It is also the first Haiku-family model to offer an adjustable effort setting. That means you can choose the balance between cost and reasoning capability per task. There is less need to handle simple classification and multi-step reasoning with the same setting. Still, the right setting has to be decided through real inputs and quality evaluation.
The wording about speed also has to be read with its conditions. Anthropic describes Haiku 5.5 as its fastest model on a standard speed basis for each model. A footnote adds that it is slower than Opus models running in Fast Mode.
Token Pricing: Two Tiers Around a 100K-Token Prompt
The official announcement’s price table splits Haiku 5.5 into prompts of 100,000 tokens or fewer, and above that. All amounts below are US dollars (USD) per million tokens.
| Item | Haiku 5.5, up to 100K | Haiku 5.5, over 100K | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|---|
| Cache read | $0.01 | $0.05 | $0.10 | $0.10 |
| Cache write | $0.125 | $0.625 | $1.25 | $2.50 |
| Input tokens | $0.10 | $0.50 | $1.00 | $2.00 |
| Output tokens | $0.50 | $2.50 | $5.00 | $10.00 |
The 100K tokens in the table is a price tier based on prompt length. When calculating call costs, you need to look not only at input and output but also at cache read and write usage separately.
As an example, suppose a request that doesn’t use the cache has 20,000 input tokens and 2,000 output tokens. For Haiku 5.5 with a prompt of 100K tokens or fewer, the pure input and output token cost is as follows.
Input: 20,000 / 1,000,000 × $0.10 = $0.002
Output: 2,000 / 1,000,000 × $0.50 = $0.001
Total: $0.003
Plugging the same token counts into Haiku 4.5’s rates gives $0.03. Under this assumption, the token cost drops by 90%. This example excludes separate costs such as tool use, and it doesn’t mean the actual token counts would be the same if you gave the same text or task to both models.
How the 75% Average Cost Drop Differs from the 90% Per-Token Cut
Anthropic says Haiku 5.5’s average operating cost is about 75% lower than Haiku 4.5’s. You shouldn’t apply this figure as a discount rate on every request.
According to the announcement footnote, the rates dropped 90% for prompts of 100K tokens or fewer, and 50% above that. About 90% of requests to the previous Haiku model were 100K tokens or fewer, and the average operating cost calculation also reflects the change in per-task token usage due to the tokenizer change. The explanation is that Haiku 5.5 may use somewhat more tokens for the same task.
When setting an actual budget, it is more accurate to measure input, output, and cache usage along with retry counts on representative tasks, rather than multiplying your existing bill by the average savings rate.
Benchmarks: A Wider Set of Candidates for Small Tasks
Here are the main figures published in the official announcement. Knowledge-work scores and accuracy rates are on different scales, so don’t compare them against the same standard.
| Evaluation | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 (reference) |
|---|---|---|---|---|
| GDPval-AA v2.1, knowledge work | 1620 | 735 | 1437 | 1840 |
| AA-Briefcase v1.1, knowledge work | 1578 | 614 | 1336 | 1824 |
| OSWorld 2.1, offline subset | 72.4% | 15.7% | 48.9% | 83.9% |
| Humanity’s Last Exam, no tools | 45.9% | 10.2% | — | 56.9% |
| Humanity’s Last Exam, with tools | 57.4% | 18.7% | — | 64.5% |
| Terminal-Bench 4.0, agentic coding | 39.2% | 0.0% | 16.4% | 70.6% |
| FrontierCode 1.1 Main, agentic coding | 46.4% | — | 42.4% | 52.1% (Xhigh) |
| Chartography, visual reasoning without tools | 46.4% | 6.4% | 29.1% | 61.6% |
— marks an item with no value shown in the official table. The OSWorld result is an offline subset, and the Sonnet 5.5 entry for FrontierCode is marked with the Xhigh setting. You need to check the tools and settings of each evaluation, and you can’t generalize an advantage on one benchmark to an advantage on all work. The detailed source for the evaluation methods is the Haiku 5.5 system card.
From Toris’s point of view, what stands out first are the knowledge-work and computer-use metrics. That is because the set of candidates for tasks like reading documents and picking out needed information, or handling a fixed screen flow, could widen.
The coding metrics need to be read alongside. On Terminal-Bench 4.0, Haiku 5.5 scores 39.2% and Sonnet 5.5 scores 70.6%. The official announcement also says Sonnet 5.5 and Opus 5.5 are the better choice for complex, multi-step agentic coding. There is reason to evaluate separately the model that owns a whole feature implementation and the model that owns subtasks like file exploration and summarization.
Monthly API Credits for Max and Team Subscribers
This announcement also includes monthly API credits for use on the Claude Platform.
| Subscription | Monthly API credit in the official announcement |
|---|---|
| Claude Max 5x | $100 |
| Claude Max 20x | $200 |
| Claude Team | Up to $500, shared among users in the team |
Anthropic says it will roll out in stages during launch week. The Team maximum of $500 is not an amount each user receives, but credit shared by all users in the team. It also says the benefit isn’t Haiku-only and can be used with any model on the Claude Platform.
The purpose of this credit is to let you experiment with tools, apps, and agents that call the API. The subscription app’s usage limits and API credits should be kept separate. Whether it applies to your actual account and the detailed terms have to be checked in the official Help Center article linked from the announcement. The table above shows the amounts stated in the announcement text, and I did not calculate on assumptions about separate conditions such as rollover or expiration.
Sonnet 5.5’s Cache Reads Are Also Half Price
Sonnet 5.5’s cache read price was cut from $0.20 to $0.10 per million tokens. Anthropic says cache reads account for a large share of token consumption, so Sonnet 5.5’s cost drops by about 20% for most agent work.
A 50% cut in the cache read rate and a roughly 20% drop in overall task cost are different figures. The actual savings also vary with how much you use the cache. For a coding flow that repeatedly uses the same system instructions and repository context, it is worth checking the cache usage of your existing Sonnet calls along with adopting Haiku.
Where to Start Testing in Product Development
For a one-person studio, the first thing to consider is not swapping models all at once but picking work with high repetition volume and clear completion conditions. What follows is an adoption suggestion based on the official announcement, not an operating case I have already applied and measured.
| Task to hand off | What the small model returns | What to check |
|---|---|---|
| Repository exploration | Relevant file paths and the code locations that support them | Does it match the actual files and code? |
| Classifying customer inquiries | A category, a summary, and the issue to pass to the owner | Does it avoid misclassifying important inquiries? |
| Summarizing documents and conversations | Decisions, open items, next actions | Are dates, figures, and sources preserved? |
| Coding assistant research | Candidate edits and expected impact | Is there enough evidence for a larger model or a human to review? |
For example, while a larger model carries out a feature implementation, Haiku can gather evidence from related files or documents. If you set the result format and completion conditions for the subtask in advance, it is easier to evaluate whether the fast response actually turns into saved working time.
It is good to put accuracy, completion time, call cost, and retry cost into the evaluation together. Even with low token rates, if you call the larger model more often to fix wrong results, the advantage shrinks. For tasks that touch databases or browsers, you also have to design access scope and result verification along with the model choice.
This view connects with the harness engineering post I wrote earlier. You need a structure that assigns work suited to the model, supplies the necessary context and tools, and checks results, for price changes to turn into changes in a product’s operating cost.
My Takeaway from This Announcement
The change worth noting in Haiku 5.5 is that the economic room to place a small model on repetitive work has grown. The low rates in the 100K-and-under tier, the adjustable effort, and the subagent use were presented together. Add the Sonnet cache read cut and subscribers’ API credits, and there are more grounds to start small experiments.
If you are considering adoption, you can start by picking a few representative tasks and comparing them against your existing model with the same inputs and completion criteria. The first step is checking how much completion time and total cost change within the range where quality holds. For complex coding work, it is worth first looking at whether you can keep the larger model’s strengths while handing repetitive subtasks to Haiku.
References
- Anthropic: Introducing Claude Haiku 5.5: The source for the announcement date, uses, performance table, token pricing, API credits, and cost calculation footnotes.
- Arca Live: Introducing Claude Haiku 5.5: The Korean introduction and translation that prompted this post. Prices and figures were cross-checked against Anthropic’s original.
Related posts: Claude Code Trends · AI Coding Agents 2026 Overview
Advertisement