Claude Haiku 5.5: Pricing, Benchmarks, and Subagent Setup

What Is Claude Haiku 5.5?

The Haiku 5.5 model page lists Claude Haiku 5.5 for high-volume, latency-sensitive tasks such as classification, extraction, and routing. Its comparison table puts Haiku 5.5 below Sonnet 5.5 and Opus 5.5 on price and rates its latency the fastest in the current Claude lineup. Its specifications:

• Model ID - claude-haiku-5-5 on the Claude API, Google Cloud, Microsoft Foundry, and Claude Platform on AWS, and anthropic.claude-haiku-5-5 on Amazon Bedrock.
• Context and output - a 1M-token context window and 128K max output tokens, or up to 300K output tokens on the Message Batches API with the output-300k-2026-03-24 beta header.
• Thinking and effort - adaptive thinking is on by default, the effort parameter sets how much the model thinks, and the default effort is medium.
• Inputs and outputs - text and images in, text out, with a reliable knowledge cutoff of June 2026.
• Tokenizer - the same tokenizer as Claude 4.7 and later models, which counts the same text as more tokens than Haiku 4.5 did.
• Retirement - not sooner than October 7, 2027.

The Claude Opus 5.5 breakdown covers the Opus-tier model in the same family, including the API changes it introduced in September.

Claude Haiku 5.5 Pricing

Anthropic's launch post prices Haiku 5.5 in two tiers split at a 100K-token prompt, a size it says covers most requests to the previous Haiku model. Prices per million tokens:

• The 75% average - per the launch post's footnote, Haiku 5.5 is priced 90% lower than Haiku 4.5 on requests up to 100K tokens and 50% lower above that. On Haiku 4.5, 90% of requests fell under 100K tokens, and the figure also accounts for the newer tokenizer using slightly more tokens per task.
• Batch and caching discounts - the model page lists a 50% discount on Batch API input and output, and 1-hour cache writes at $0.20 per million tokens on prompts up to 100K.
• Sonnet 5.5 cache reads - at the same launch, Anthropic halved Sonnet 5.5's cache read price from $0.20 to $0.10 per million tokens, which it says makes Sonnet 5.5 around 20% cheaper on most agentic work.
• API credits for subscribers - Claude Max 5x subscribers get $100 a month in Claude Platform API credits, Max 20x subscribers get $200, and Team subscribers get up to $500 pooled across users, usable on any Claude model.

Recount prompts before you rebudget. Anthropic's Haiku 5.5 migration guide says the same input text produces approximately 30% more tokens on Haiku 5.5 than on Haiku 4.5, so usage fields, token counts, max_tokens limits, and cost estimates measured on Haiku 4.5 all need redoing with model set to claude-haiku-5-5.

Claude Haiku 5.5 Benchmarks

Anthropic's launch post scores Haiku 5.5 against Haiku 4.5, OpenAI's GPT-6 Luna, and Sonnet 5.5, which it includes for reference. The Haiku 5.5 system card documents how each evaluation ran.

Haiku 5.5 is the first Haiku-class model with an adjustable effort setting, so one model ID can trade cost for intelligence per request. The launch post charts OSWorld, GDPval-AA, and Humanity's Last Exam at each effort level from low to max, with cost per attempt on the horizontal axis. The scores point to where the model fits:

• Computer use - Haiku 5.5 scores 72.4% on the OSWorld 2.1 offline subset against 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna, and the launch post names browser use and live customer support as speed-sensitive tasks the model suits.
• Complex agentic coding - Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 against 39.2% for Haiku 5.5. Anthropic says Sonnet 5.5 and Opus 5.5 remain the better choices there, and that Haiku 5.5 suits narrowly scoped jobs such as compaction, summarization, and subagent work.
• Speed - Haiku 5.5 is Anthropic's fastest model at each model's standard speed, though it runs slower than the Opus models in Fast Mode.

Early-access customers quoted in the launch post reported these results:

• Asana - over a 30% reduction in task-completion latency and up to 2.5x faster inference per agent turn on its AI Teammates eval suite, compared with the model it uses today.
• HubSpot - 92.8% averaged over three runs on its simulated-portal CRM suite, the best score HubSpot has seen there, plus the fastest completion and lowest false positive rate on a stale-records audit task.
• AlphaSense - 0.84 against 0.76 for Haiku 4.5 across 400 queries to Ask in Document, a feature that handles about 8M calls a week.
• Box - 11 points higher than Haiku 4.5 at about half the latency in early testing.

Claude Haiku 5.5 as a Subagent in Claude Code

Claude Code needs v2.1.293 or later for Haiku 5.5. Its model configuration docs say the haiku alias resolves to Haiku 5.5 on the Anthropic API, but still resolves to Haiku 4.5 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry, and Claude Platform on AWS. To run it as the main model, use /model claude-haiku-5-5 or start Claude Code with claude --model claude-haiku-5-5.

The more common setup keeps Opus 5.5 or Sonnet 5.5 as the lead and hands narrow jobs to Haiku 5.5. A subagent file in .claude/agents/ sets the model and effort in its frontmatter:

---
name: test-log-summarizer
description: Summarizes test runner and CI output into failing tests, error lines, and likely flaky tests. Use after a test run finishes.
tools: Read, Grep, Glob
model: haiku
effort: low
---
You summarize test output. For each failing test, report the file, the
assertion that failed, and the first error line. Flag tests that failed
and then passed on retry as likely flaky. Do not edit any files.

Claude Code picks a subagent's model in this order:

• The model parameter Claude passes for that specific invocation.
• The model field in the subagent's frontmatter, where inherit selects the main conversation's model.
• The CLAUDE_CODE_SUBAGENT_MODEL environment variable, which also sets the default for agent team teammates and workflow agents.
• The main conversation's model.

The Claude Code subagents guide covers the remaining frontmatter fields and where subagent files live. Alex Wang of Rogo described the lead-and-subagent split in the launch post:

Anthropic's Prompting Claude Haiku 5.5 guide flags a subagent risk: at low and medium effort, the model sometimes reports a code change as done without running a check that exercises it. The guide supplies a system-prompt paragraph that tells coding agents to run the project's tests, type-checker, or build before reporting.

To check whether a Haiku 5.5 subagent did what it reported, Agent Assurance from TestMu AI tests how agents behave across workflows, tools, and actions. Its rook CLI derives a suite from your agent's code or spec, invokes the agent for real, and checks each criterion against what the run changed, such as files on disk and tool calls, instead of the agent's own summary. Criteria it could not verify are reported separately, so an unverified "done" never counts as a pass.

For a model switch, rook finds Claude Code agents defined in .claude/agents/, and a command profile such as claude -p "{{goal}}" invokes the agent the way a user would. Every run is diffed against the last, so scenarios that newly fail after you set model: haiku are reported apart from flaky ones. Agent Assurance is in pre-alpha, and rook installs with npm install -g @testmuai/rook.

Anthropic's Browser and Computer Use Tools on Claude Haiku 5.5

Haiku 5.5 supports Anthropic's browser use tool on the Claude API and Google Cloud. One browser_toolset_20260801 entry in tools gives Claude 27 member tools by default, such as navigate, read_page, left_click, and screenshot, and your application executes every call against its own browser automation. Nothing runs on Anthropic's side.

• New to the Haiku line - the migration guide notes that Haiku 4.5 does not support the browser use tool.
• Platform coverage - the browser use tool is not available on Amazon Bedrock, Claude Platform on AWS, Microsoft Foundry, or in Claude Managed Agents.
• Computer use toolset - on the Claude API and Google Cloud, Haiku 5.5 takes computer use only through computer_toolset_20260801, and a request that declares computer_20250124 returns a 400 error.
• SDK support - Anthropic's launch post says the Python and TypeScript SDKs now support computer use and browser use in beta, and calls Haiku 5.5 especially well suited to both.
• Untrusted pages - the tool docs recommend a domain allowlist enforced at the network layer and re-checked after redirects, and page reads built from the accessibility tree or visible text rather than raw DOM, so hidden text never reaches Claude.

Your application has to supply that browser. TestMu AI Browser Cloud provides real Chrome sessions on demand for AI agents, with Playwright, Puppeteer, and Selenium adapters, a built-in tunnel to localhost and staging, and session transparency for debugging a failed run. The what is Browser Cloud doc covers session setup, and browser infrastructure for AI agents walks through scaling, debugging, and deploying those sessions.

To confirm what a browser agent changed in your web app, Kane CLI runs a natural-language objective in a real Chrome browser and returns a pass or fail with a replayable .evidence pack of per-step screenshots, DOM snapshots, and console and network logs. It attaches to an existing Chrome with --cdp-endpoint or to a remote browser with --ws-endpoint, and it installs with npm install -g @testmuai/kane-cli.

Breaking Changes From Claude Haiku 4.5

The migration guide lists these request changes for code that called Haiku 4.5:

• Automated migration - in Claude Code, /claude-api migrate this project to claude-haiku-5-5 runs the bundled Claude API skill, which swaps the model ID, applies the breaking changes, and returns a checklist of items to verify by hand.
• Thinking display - Haiku 5.5 returns each thinking block with an empty thinking field by default. Set display to summarized to receive summarized thinking, as Haiku 4.5 returned.
• Turning thinking off - on the API, thinking: {"type": "disabled"} works at low, medium, and high effort and returns a 400 error at xhigh and max. In Claude Code, thinking can't be turned off on Haiku 5.5.
• Priority Tier - not supported on Haiku 5.5, so organizations with a Haiku 4.5 Priority Tier commitment need to plan capacity separately.
• Stored conversations - Haiku 5.5 thinking blocks work only in the account that produced them or a linked account. Replayed through any other account, they are dropped before the model sees them, and the request still succeeds.

How to Test Agents Before You Switch to Claude Haiku 5.5

Anthropic expects existing Haiku 4.5 prompts to perform well without changes, and its prompting guide still documents several places where Haiku 5.5 behaves differently. Before moving production traffic, run your AI agent testing suite against these behaviors:

• Compare effort levels - run the same evals at two or three levels. The prompting guide says that in long agent prompts, low is more likely to skip a search, stop early, or skip a check, and that moving from low to medium roughly halved early stopping while more than doubling output tokens per attempt.
• Catch unverified completions - for coding agents, compare what the agent reports against the tests, builds, and files the run actually touched.
• Check for empty replies - at xhigh effort in multi-turn chats, the model sometimes writes its whole answer in its thinking and ends the turn with no visible text.
• Test search behavior - give the model today's date when it has a search tool. The guide reports that the date grounded answers in recent results, and that a blanket instruction to always search triggered searches on half of the prompts that needed none.
• Pressure-test chatbot rules - Anthropic suggests a system-prompt line that keeps rules in force when a user argues, gives a sympathetic reason, claims an approved exception, or keeps asking. Script those exact patterns as test conversations.
• Route mid-turn messages correctly - Haiku 5.5 is trained to resist prompt injection through tool results, so user text placed inside a tool_result block can be ignored as untrusted. Deliver it as a user turn instead.
• Handle refusals in the client - Haiku 5.5 has no server-side fallback, and resending a declined request usually returns another refusal.

For chat and voice agents moving to Haiku 5.5, Agent Testing from TestMu AI generates 60 to 100+ scenarios from uploaded docs and PRDs or connected Jira and Confluence, covering edge cases, adversarial inputs, and persona-specific conversations. It scores each conversation on 9 quality metrics, including hallucination, completeness, and context awareness, and returns a Green, Yellow, or Red production-readiness verdict. One suite run on Haiku 4.5 and again on Haiku 5.5 shows whether the cheaper model holds the same quality bar.

Note: Test the chatbots and voice agents you move to Claude Haiku 5.5 with TestMu AI Agent Testing. Try TestMu AI free!

Safeguards and Refusals on Claude Haiku 5.5

Anthropic's launch post reports far fewer instances of misaligned behavior than on Haiku 4.5, and a lower willingness to cooperate with misuse. The safeguards also change what an agent can be asked to do:

• Cybersecurity - more restrictive than Haiku 4.5's safeguards but less restrictive than those on other recent models. They permit a wider range of defensive tasks than Sonnet 5.5's safeguards and still block penetration testing.
• Biology - the same safeguards as Sonnet 5, Sonnet 5.5, and Opus 5, which allow research biology questions and restrict requests judged likely to cause harm.
• Verification programs - organizations doing wider-ranging work can apply to the Life Sciences Verification Program and the Cyber Verification Program.
• Refusal categories - a declined request returns stop_reason set to refusal, with stop_details.category set to cyber, frontier_llm, bio, or general_harms. Benign work can also trigger cyber and general_harms.

A refusal arrives as a regular response rather than an error, so a pipeline that reads content without checking the stop reason can pass an empty result downstream. Teams that run security scans through an agent should test that path directly, and prompt injection testing covers the attack side of the same tool-result boundary.

Getting Started With Claude Haiku 5.5

Start with one high-volume job, such as summarizing test logs or classifying support tickets. Switch its model to claude-haiku-5-5, remove sampling parameters and prefill, and run your existing evals at low and medium effort next to the same evals on Haiku 4.5.

For a chat or voice agent, Agent Testing scores both runs on the same metrics, and the getting started with Agent Testing guide covers connecting your agent.