A short subagent turn of Claude Code costs around one-tenth of a cent with Claude Haiku 5.5 and nearly two cents with Sonnet 5.5. In a long turn of the main session, that gap narrows from about 17 times to about 3. In that difference lies the whole decision: Haiku 5.5 is a model for subagents, and using it as a cheaper main model wastes almost all the savings.
What is Claude Haiku 5.5?
Claude Haiku 5.5 is Anthropic’s small and fast model, released on October 7, 2026. Anthropic positions it for high-volume and cost-sensitive work: summaries, compression, classification, browser use, and most importantly here, functioning as a subagent of Opus 5.5 or Sonnet 5.5 in coding tasks.
The model ID is claude-haiku-5-5. It has a 1M token context window, maximum output of 128K, and knowledge through June 2026. It’s the first Haiku with adjustable effort level, and defaults to medium. According to Anthropic, it’s available across all its platforms, including AWS, Google Cloud, and Microsoft Azure.
Anthropic is unusually direct about its ceiling. On Terminal-Bench 4.0, its agentic coding benchmark, Haiku 5.5 scores 39.2% versus Sonnet 5.5’s 70.6%, and the announcement says Sonnet and Opus “remain better choices for complex agentic coding”. All benchmark figures are reported by Anthropic.
| Benchmark | Haiku 5.5 | Sonnet 5.5 | Haiku 4.5 |
|---|---|---|---|
| Terminal-Bench 4.0 | 39.2% | 70.6% | 0.0% |
| FrontierCode 1.1 (Main) | 46.4% | 52.1% (Xhigh) | — |
| OSWorld 2.1 (offline subset) | 72.4% | 83.9% | 15.7% |
What is the price of Claude Haiku 5.5 compared to Sonnet 5.5 and Opus 5.5?
With short prompts, Haiku 5.5’s input costs one-twentieth of Sonnet 5.5’s. Anthropic API list prices, per million tokens (claude haiku 5.5 pricing):
| Haiku 5.5 (prompt ≤100K) | Haiku 5.5 (prompt >100K) | Sonnet 5.5 | Opus 5.5 | |
|---|---|---|---|---|
| Input | $0.10 | $0.50 | $2 | $4 |
| Output | $0.50 | $2.50 | $10 | $20 |
| Cache read | $0.01 | $0.05 | $0.10 | $0.20 |
| Cache write (5 min) | $0.125 | $0.625 | $2.50 | $5 |
Two details weigh more than the headline.
The 100K threshold. Haiku 5.5 is the only current Claude model priced by prompt length. Above 100K tokens, each line costs five times more. Anthropic says prompts up to 100K were about 90% of requests to the previous Haiku, and that’s where their “around 75% less” on average comes from. That average is from Anthropic and already includes that Haiku 5.5’s new tokenizer uses somewhat more tokens per task. It’s not a savings measured over your workload.
Sonnet 5.5 also dropped the same day. Anthropic cut Sonnet 5.5’s cache read in half, from $0.20 to $0.10. Opus 5.5 stays at $0.20. As of October 8, 2026, those are list prices. This series has already had price changes ten days apart, so check them before budgeting on them.
How much does Claude Code cost per turn with Haiku, Sonnet, or Opus?
It depends almost entirely on how much context the turn loads. Two illustrative turns, with my arithmetic on list prices and not counting cache writes.
A short subagent turn: 30K tokens read from cache, 3K new input and 1K output. The prompt is 33K, so Haiku stays in its cheap tier.
| Haiku 5.5 | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| ~$0.0011 | ~$0.019 | ~$0.038 |
A long main session turn: 100K in cache, 5K new input and 2K output. The prompt is 105K, so Haiku moves to the expensive tier.
| Haiku 5.5 | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| ~$0.0125 | ~$0.040 | ~$0.080 |
Haiku is about 17 times cheaper than Sonnet on the first turn and about 3 times on the second. My reading is that cached tokens count toward the 100K threshold, because Anthropic defines pricing by prompt length. Check consumption for the first week before relying on that.
That’s why the subagent approach holds. A Claude Code subagent starts with fresh context: its own system prompt, the task Claude delegates to it, your CLAUDE.md files, and a snapshot of git status. It doesn’t carry your conversation history, and the built-in Explore agent doesn’t even load CLAUDE.md or git status. Subagent prompts start small and stay in the up-to-100K tier, exactly where Haiku’s price advantage is greatest.
When is Claude Haiku 5.5 worth it instead of Sonnet 5.5?
In the Haiku vs Sonnet comparison, Haiku 5.5 wins when the task is scoped, especially read-heavy and easy to verify. Sonnet 5.5 or Opus 5.5 are still the choice when the task requires judgment.
- Haiku 5.5 territory: code search and exploration, summaries of logs or test output, running the test suite and reporting only failures, extracting data from a large file, classifying or triaging issues. Anthropic’s own examples have that shape: compression, summaries, and subagent work.
- Sonnet 5.5 or Opus 5.5 territory: anything that writes non-trivial code, debugging without a known reproduction, architectural decisions, and changes across multiple files. A 39% on Terminal-Bench is a clear signal not to hand the keyboard to Haiku.
Customer testimonials point the same direction, though Anthropic published them. Cognition says Haiku 5.5 is already an option companion model in Devin Fusion, with Opus 5.5 as the main model. Rogo describes a larger model putting together the deliverable while a Haiku subagent pulls a single line from a 10-K. The pattern: the main model decides and Haiku searches.
How do you use Haiku 5.5 in Claude Code subagents?
First, update. Claude Code documentation asks to use v2.1.293 or higher with Haiku 5.5, which is also the version where the haiku alias starts pointing to Haiku 5.5 in the Anthropic API.
claude update
There are three ways to send work to it, from most scoped to most broad. If Claude Code subagents are new to you, start with our guide to sub-agents in Claude Code.
1. Per subagent. It’s the option I’d start with. Define model: haiku in the frontmatter of your own subagent. Subagent files go in .claude/agents/ (per project, to version them with the repo) or in ~/.claude/agents/ (across all your projects):
---
name: log-summarizer
description: Summarizes test and build output, reporting only failures. Use proactively after running tests.
tools: Read, Grep, Glob, Bash
model: haiku
---
Summarize the output you are given. Report only failing tests with their error messages.
The same frontmatter accepts an effort field if you want Haiku to work below or above its medium default.
2. Replace the built-in Explore agent. By default, Explore uses your main conversation model, so in a session with Opus your code searches run on Opus. According to the documentation, a user or project subagent called Explore replaces the built-in one, keeps its own model field, and you can define it with model: haiku. Your definition also replaces the built-in system prompt, so write one and give it read-only tools.
3. A default model for all subagents. CLAUDE_CODE_SUBAGENT_MODEL sets the model for subagents that don’t have one assigned another way. If you combine it with CLAUDE_CODE_SUBAGENT_MODEL_FORCE (v2.1.257 or higher), it’s imposed on all, Explore included:
{
"env": {
"CLAUDE_CODE_SUBAGENT_MODEL": "haiku",
"CLAUDE_CODE_SUBAGENT_MODEL_FORCE": "1"
}
}
I wouldn’t use the forced version. It also puts Haiku on the general-purpose subagent, and that agent edits code.
Run /tasks while a subagent is working to confirm what model it’s actually using. If you want to try Haiku as the main model, /model claude-haiku-5-5 works, but expect the long-session price from the earlier table.
On a Pro or Max plan, subagent requests count toward the same usage limits as your main conversation. The dollar figures in this article are API list prices. At publication time, I found no Anthropic documentation on how much each model consumes from plan limits.
Does the haiku alias point to Haiku 5.5 on Bedrock, Google Cloud, or Azure?
No, not yet. As of October 8, 2026, Claude Code documentation assigns haiku to Haiku 5.5 only on the Anthropic API. On Amazon Bedrock, Google Cloud’s Agent Platform, Microsoft Foundry, and Claude Platform on AWS, it still points to Haiku 4.5. A subagent with model: haiku on those providers runs the earlier model at the earlier price.
There, set the model explicitly with ANTHROPIC_DEFAULT_HAIKU_MODEL, using your provider’s Haiku 5.5 ID. The same variable controls the model Claude Code uses for its background tasks.
What changes for a technical team?
Haiku 5.5 doesn’t cheapen your main model. What it does is make delegating cheap enough that the question shifts from “what model do we use?” to “what work goes to the main model?”. That’s a routing policy, and routing policies go in version control. A .claude/agents/ directory with a few Haiku subagents for search, summaries, and test triage is a policy everyone inherits with a git pull.
The top of the stack doesn’t change: Opus for ambiguous work, Sonnet for well-scoped implementation. Haiku adds a third tier below, for work that never justified the tokens of a frontier model.
A related change: Anthropic says it’s starting this week to deliver a monthly API credit to Claude Max and Team subscribers, $100 on Max 5x, $200 on Max 20x, and up to $500 shared on Team. That’s enough to measure these numbers over your own workloads before setting a policy.
You? What Claude Code task would you first move to Haiku 5.5?
- Search and exploration subagents
- Summaries of tests and logs
- Nowhere, I’m staying with Sonnet or Opus
- I prefer another option (tell us which)
Comment below or, if this article reached you by email, reply directly to the email: your reply is published here.
Related: Claude Sonnet 5.5 vs Opus 5.5: in Claude Code, choosing a model is already a cost decision