Claude Code Haiku: model selection, costs, and focused workflows

What Haiku contributes to Claude Code

Claude Code is the agent application: it reads context, calls permitted tools, and works through the assigned task. Haiku is one model it can use. A direct Haiku API call is not automatically a Claude Code session, and neither setup gains access to private systems without the appropriate context and credentials.

As checked on September 9, 2026, Anthropic lists Haiku 4.5 as the current Haiku model. Its published specifications include a 200K-token context window, up to 64K output tokens, text and image input, and text output. Those are model limits, not a promise that every Claude Code session has that much free space or will produce a correct answer at the limit.

Anthropic positions Haiku for lower-latency work. Its launch report includes coding evaluations, but an evaluation result is not evidence about your repository. Do not turn it into a claim that all fixes are instant, that no review is needed, or that Claude Code becomes an inline-completion product when Haiku is selected.

Select a model and verify what actually runs

For a new terminal session, the documented selection is:

claude --model haiku

This assumes Claude Code is installed and authenticated. The flag selects the model for that launch. In an existing interactive session, /model haiku switches models; current Claude Code also saves that selection as a user default. To change only the current session through the picker, open /model and use its session-only choice. See the model configuration reference for current precedence and provider differences.

The convenient haiku alias should not be treated as an immutable model version. For a reproducible comparison, record the resolved model, provider, Claude Code version, prompt, and relevant settings. On the Claude API, Haiku 4.5's pinned ID is claude-haiku-4-5-20251001; third-party providers use their own deployment or version identifiers. Check access before substituting an ID into a script.

Model selection does not widen permissions. If a session cannot read a file or use a tool, choosing a different model does not authorize it. Conversely, a smaller model with broad shell or connector access can still make consequential changes.

Give Haiku a job with a clear finish line

Good evaluation tasks have both a small input and an observable result. These are proposed uses, not test results from this article:

Give the model enough evidence to do the job, but not a whole repository dump when two files suffice. If the result misses a dependency, repeatedly contradicts the source, or needs substantial correction, broaden the investigation or use another model. Measure the completed job, including retries and review, rather than declaring the cheapest token rate the cheapest workflow.

Using Haiku in a subagent

Claude Code supports custom subagents with a model field and a tool allowlist. A document-checking subagent could select model: haiku and use only Read, Grep, and Glob when it needs to inspect already-approved local material. That subagent cannot itself run eesel CLI without a shell tool; an authorized parent session or person would first obtain the permitted material.

The current subagent reference also corrects an old assumption: built-in Explore no longer always runs on Haiku. It inherits the main model, with an Opus cap on the Claude API. Check /tasks while a subagent runs to see its actual model. If your goal is a Haiku comparison, configure and verify that choice rather than assuming delegation selected it.

Understand the cost comparison

The Claude API pricing page lists Haiku 4.5 base rates of $1 per million input tokens and $5 per million output tokens. For illustration, 10,000 uncached input tokens plus 2,000 output tokens would cost $0.02 at those base rates. That is arithmetic, not a measured Claude Code task or an all-in quote; repeated turns, tool use, caching, and provider billing can change the total.

Subscription usage is a different arrangement. Do not multiply an API token estimate into a guarantee about how many jobs a Claude plan includes. Record actual usage under the account and provider you will use, and include the time spent correcting outputs. Switching models is valuable only if the finished result still meets the task's standard.

Use eesel CLI for a focused support-instruction review

Suppose a support owner wants to resolve an instruction conflict before opening a new queue. In this fictional example, one clause says to send failed verification cases to the identity team, while another says all unresolved login questions go to general support. The owner needs the conflict identified and a precise proposal—not a model deciding which team owns the case.

A person can run eesel CLI directly, or ask Claude Code, Codex, or Cursor to run the approved observations and read the JSON output. Node.js 18.17 or newer is required for the documented npx form. Start with an existing authorized workspace:

npx @eesel/cli login
npx @eesel/cli whoami
npx @eesel/cli agents

Verify the intended workspace, then replace the placeholder below with the real teammate ID or name chosen by its owner. For an unattended script, EESEL_API_URL and EESEL_API_TOKEN override stored credentials; verify the target rather than assuming a previous login still controls it.

INSTRUCTION_REVIEW_AGENT="REPLACE_WITH_OWNER_APPROVED_AGENT"
npx @eesel/cli status --agent "$INSTRUCTION_REVIEW_AGENT"
npx @eesel/cli instructions --agent "$INSTRUCTION_REVIEW_AGENT"

These observations read the current setup. The coding agent can compare the returned instructions with an owner-provided escalation policy and produce a small report: exact conflicting clauses, the cases affected, and the decision still needed. Share only the material authorized for that coding session. The model should not invent a policy document or infer that it checked customer identity records.

For this example, ask: “Identify the overlap for a failed verification case. Quote the relevant clauses. Do not choose a destination or edit the teammate. Ask the owner which rule takes precedence.” That is a focused extraction and comparison job suited to evaluating a smaller model. If the clauses have exceptions elsewhere, the report should request those sources rather than quietly resolving the gap.

After the owner decides, the coding agent can draft the narrow instruction change. Inspect npx @eesel/cli instructions --help for the supported edit syntax, preview the proposed write with --dry-run, and apply only the approved change. Read instructions back to confirm what was saved. A preview shows the proposed request, not how future conversations will behave.

For response evaluation, the owner selects a separate non-production teammate and sets consequential connected actions to Disabled in Actions and Approvals. Test a failed-verification question and an ordinary unresolved-login question in separate fresh conversations using the approved policy. Check the actual replies, activity, and held approvals. No identity check or transfer should be claimed merely because a response names a team. Setup observations are free; these eesel chats are billed and need an approved budget.

Selecting Haiku for the coding session does not select the model inside eesel. The practical connection is the workflow: the coding agent helps the owner inspect and improve the existing teammate through documented commands.

Start with a small, reviewable eesel CLI task

Use Haiku where a bounded job and a clear check make its tradeoffs visible. For support operations, start with one instruction conflict and its approved source—not an instruction to rewrite the whole workspace. The person reviewing the result keeps the policy decision, and the dashboard and CLI show the same teammate configuration.

Try eesel to inspect your support setup and turn an approved finding into a tested teammate improvement.