AI Coding Tools
Claude Haiku 5.5 as a subagent: what to move, what to keep
On this page
- What Haiku 5.5 costs, and where the price line sits
- Why a subagent loop crosses 100K sooner than you expect
- One exploration subagent, three models
- Which subagent jobs to move to Haiku 5.5
- Setting up Haiku 5.5 subagents in Claude Code
- Haiku 5.5 in GitHub Copilot
- Bottom line
- Frequently Asked Questions
Claude Haiku 5.5 shipped on October 7 at $0.10 per million input tokens, a tenth of what Haiku 4.5 charged, and Anthropic pitches it as a subagent for Sonnet 5.5 and Opus 5.5. The price is real. What most coverage skips is that Claude Code will not route anything to it unless you tell it to, and on four of the five platforms Anthropic sells it through, the obvious setting gets you the old Haiku instead.
Here is which Claude Haiku 5.5 subagent work makes sense, how to set it up so it actually runs there, and where the 100,000-token price step changes the math.
What Haiku 5.5 costs, and where the price line sits
Rates per million tokens from Anthropic’s pricing page, checked October 14, 2026:
| Model | Input | Output | 5-min cache write | Cache read |
|---|---|---|---|---|
| Haiku 5.5, prompt up to 100K | $0.10 | $0.50 | $0.125 | $0.01 |
| Haiku 5.5, prompt over 100K | $0.50 | $2.50 | $0.625 | $0.05 |
| Sonnet 5.5 | $2.00 | $10.00 | $2.50 | $0.10 |
| Opus 5.5 | $4.00 | $20.00 | $5.00 | $0.20 |
Two details decide whether you get the cheap row. First, the step applies to the whole request, not the tokens above the line. Second, and this is the question several launch write-ups left open, cached tokens count. Anthropic’s pricing page is explicit: a request’s prompt length “counts all of its input tokens, including cache reads and cache writes,” and a request over the threshold “pays the higher prices even when part of its prompt is a cache hit.”
These are API and Console prices. On a Claude subscription plan, Claude Code doesn’t bill you per token, so the dollar figures below apply to pay-as-you-go accounts.
Why a subagent loop crosses 100K sooner than you expect
A subagent in Claude Code is an agent loop. Every turn resends the whole conversation so far: system prompt, tool definitions, every file it has read, every tool result. The prompt grows each turn, and the turn that crosses 100,000 tokens and every turn after it pays the top row.
Two changes in Haiku 5.5 push that crossing earlier. Anthropic’s notes on what’s new in Haiku 5.5 say the same text counts as roughly 30% more tokens than on Haiku 4.5, because 5.5 uses the newer tokenizer, the same change that inflated Sonnet 5 token counts. And through the API, Haiku 5.5 keeps thinking blocks from every earlier assistant turn in context as input tokens, where Haiku 4.5 kept only the latest. Anthropic’s own wording: multi-turn conversations “carry more input tokens than the tokenizer change alone explains.”
Thinking is on by default, and in Claude Code you can’t turn it off for Haiku 5.5. The lever you do have is effort, which defaults to medium.
One exploration subagent, three models
To see what the step does in practice, take an illustrative exploration run. These are assumptions, not measurements: 30 turns, the prompt starts at 20,000 tokens and grows by 2,500 per turn, each turn writes 800 output tokens including thinking, the previous turn’s prompt is served from cache, and the new tokens are written to the 5-minute cache.
| Run | Opus 5.5 | Sonnet 5.5 | Haiku 5.5 |
|---|---|---|---|
| Grows 2,500/turn, ends at 92.5K | $1.26 | $0.63 | $0.04 |
| Grows 4,000/turn, ends at 136K (last 9 turns over the line) | $1.60 | $0.80 | $0.13 |
Crossing the line roughly triples the Haiku run. It still costs about a sixth of the same run on Sonnet 5.5 and a twelfth of Opus 5.5. So the 100K step is a reason to keep subagent briefs tight, not a reason to avoid Haiku. The argument against Haiku 5.5 for a given job is capability, and that is where the numbers are less flattering.
Which subagent jobs to move to Haiku 5.5
Anthropic’s launch table, which is vendor-reported, puts Haiku 5.5 at 39.2% on Terminal-Bench 4.0 against 70.6% for Sonnet 5.5. That is a large jump from Haiku 4.5’s 0.0%, and it is still a big gap. Anthropic says plainly that Sonnet 5.5 and Opus 5.5 remain the better choices for complex agentic coding. GitHub’s own early testing found Haiku 5.5 matched Sonnet 5 on many coding tasks with fewer tokens and steps, which is a comparison against the previous Sonnet, not the current one.
That gap points at a clean split. Move the work where a wrong answer is cheap to spot and cheap to redo:
- Codebase search. Finding files, symbols, call sites and config. The answer is a list of paths the parent can check in one read.
- Log and test-output triage. Reading 4,000 lines of CI output and returning the three failures that matter.
- Summaries for the parent. Condensing a long diff, an issue thread or a dependency changelog into what the main agent needs.
Keep on Sonnet 5.5: multi-file edits, debugging that needs a hypothesis tested across several files, and anything where the subagent’s output goes straight into a commit. A failed Haiku edit that the parent has to diagnose and redo costs more than the Sonnet run would have.
Documentation passes over legacy code sit in between. They are high-volume and read-only, which suits Haiku, but a workflow like documenting RPG IV with AI depends on reading fixed-format columns correctly, exactly the kind of detail a small model can get wrong. Test it on code you already understand before trusting it on code you don’t.
Setting up Haiku 5.5 subagents in Claude Code
Start with the version. Anthropic’s model docs say to use Claude Code v2.1.293 or later with Haiku 5.5. Run claude update.
The built-in Explore agent doesn’t use Haiku
This is the part that surprises people. Claude Code’s built-in Explore subagent runs on the main conversation’s model. If your session is on Opus 5.5, every codebase search Explore does is billed at $4 and $20 per million on API billing. Plan inherits the main model too. The only built-in that runs on Haiku is the claude-code-guide helper.
The docs give the fix: a project or user subagent named Explore overrides the built-in and keeps its own model field. Save this as .claude/agents/explore.md in your repo:
---
name: Explore
description: Fast read-only codebase search. Use to find files, symbols, call sites and config before planning or editing.
tools: Read, Grep, Glob
model: haiku
effort: low
---
Search the codebase and report what you find as file paths with line numbers.
Do not propose or make changes. Keep the final answer under 300 words.
The short answer limit matters as much as the model. Whatever the subagent returns lands in the parent’s context, and the parent pays its own rate to read it. Start at low effort and raise it to medium if the agent starts missing things.
The haiku alias is not Haiku 5.5 everywhere
On the Anthropic API, haiku resolves to Haiku 5.5. On Claude Platform on AWS, Amazon Bedrock, Google Cloud’s Agent Platform and Microsoft Foundry, Claude Code’s model docs list it resolving to Haiku 4.5. The file above would run your Explore agent on an older model that lists at ten times the per-token price, and nothing would tell you.
On those providers, set ANTHROPIC_DEFAULT_HAIKU_MODEL to your provider’s Haiku 5.5 model ID, which Anthropic lists on the Haiku 5.5 overview page. That changes what haiku points to for every agent at once.
CLAUDE_CODE_SUBAGENT_MODEL is a default now, not an override
If you set this variable months ago to push subagents onto a cheap model, check it. Since v2.1.251 it sits third in the resolution order: a model Claude passes for one invocation wins, then the subagent’s model frontmatter, then the variable, then the main model. It also doesn’t change Explore or Plan on its own.
To force one model onto every subagent, the docs add a second variable (v2.1.257 or later):
{
"env": {
"CLAUDE_CODE_SUBAGENT_MODEL": "haiku",
"CLAUDE_CODE_SUBAGENT_MODEL_FORCE": "1"
}
}
I would not do this. Forcing sends your editing subagents to Haiku as well, which is the work it does worst. Per-agent frontmatter is the better control.
To confirm any of this took effect, run /tasks while a subagent is working. Since v2.1.242 it names the model, and the effort level if one is set, on each subagent’s row. It takes ten seconds and it is the quickest way to catch the alias problem.
If you’d rather not write subagent briefs from scratch, the AI Prompt Library includes agent prompt templates you can adapt into the body of these files.
Haiku 5.5 in GitHub Copilot
GitHub made Haiku 5.5 generally available in Copilot on October 7 for Pro, Pro+, Max, Business and Enterprise. You pick it in the model picker in VS Code, Visual Studio, Copilot CLI, the cloud agent, JetBrains IDEs, Xcode and Eclipse, among others. GitHub bills it at provider list pricing under usage-based billing, so the same $0.10 and $0.50 logic carries over.
Business and Enterprise admins control it through the model policy. New models are on by default unless an admin has turned off the global default, so if you can’t see it, check that setting before assuming the gradual rollout hasn’t reached you. The same split applies: pick Haiku 5.5 for chat and agent sessions that only read, and keep Sonnet 5.5 for the ones that edit. Copilot’s agents can now do more than edit code, too, as our look at Copilot computer use on IBM i covers, and those are not jobs to hand a small model.
Bottom line
Move read-only subagent work to Claude Haiku 5.5 this week: codebase search, log and test triage, summaries for the parent. On a pay-as-you-go account the example runs above came out 12 to 30 times cheaper than on Opus 5.5, and even a run that crosses the 100K line stays far cheaper than Sonnet 5.5. Keep editing and debugging subagents on Sonnet 5.5. The Terminal-Bench gap is too wide to hand them over.
The setup is where people lose the saving. Override the built-in Explore agent, pin the model ID if you are on Bedrock, Google Cloud, Foundry or Claude Platform on AWS, and check /tasks once. If you are building longer hands-off pipelines on top of this, our guide to Claude Code scheduled agents covers the orchestration side.
Frequently Asked Questions
Does Claude Code use Haiku 5.5 for subagents by default?
No. The built-in Explore and Plan subagents run on your main conversation’s model, so an Opus 5.5 session explores at Opus rates. Only the claude-code-guide helper runs on Haiku. You opt in by defining a subagent with model: haiku. CLAUDE_CODE_SUBAGENT_MODEL only covers subagents nothing else assigns a model to, and does not change Explore or Plan by itself.
How much does Claude Haiku 5.5 cost?
$0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Prompts over 100,000 tokens pay five times that on every rate, $0.50 and $2.50, and the whole request moves to the higher rate.
Do cached tokens count toward Haiku 5.5’s 100K price line?
Yes. Anthropic’s pricing page says a request’s prompt length counts all of its input tokens, including cache reads and cache writes, and a request over the threshold pays the higher prices even when part of its prompt is a cache hit.
Is Haiku 5.5 good enough for agentic coding?
Not as the main agent. Anthropic reports 39.2% on Terminal-Bench 4.0 for Haiku 5.5 against 70.6% for Sonnet 5.5, and recommends Sonnet 5.5 or Opus 5.5 for complex agentic coding. It fits read-only search, triage and summary work under a stronger model.
Share
Grant M.
Developer with IBM i and full-stack experience. Covers AI tools and automation for software developers at PromptedDev, with a focus on real workflows, honest comparisons, and legacy system modernization.
Related reading