Claude Haiku 5.5: The Cheap Subagent With Effort Levels

1. The launch facts

Haiku 5.5 completes the 5.5 lineup next to Opus 5.5 and Sonnet 5.5. It is available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, and the docs list a single model ID, claude-haiku-5-5, with no date suffix and no separate alias.

Two details shape everything that follows. First, thinking is adaptive and on by default, so your output tokens now include reasoning unless you lower effort. Second, prices are tiered by prompt length: the cheap rate applies to prompts up to 100,000 tokens, and a request over that pays higher prices on every token.

2. What actually improved vs Haiku 4.5

Anthropic’s page says Haiku 5.5 is priced 90% lower than Haiku 4.5 for requests up to 100,000 tokens and 50% lower above that: $1 / $5 became $0.10 / $0.50. On vendor-reported benchmarks the capability jump is bigger than the price drop suggests: Terminal-Bench 4.0 goes from 0.0% to 39.2%, OSWorld 2.1 (offline subset) from 15.7% to 72.4%, and Humanity’s Last Exam without tools from 10.2% to 45.9%.

The API changed too. Haiku 4.5 had no effort parameter and used a manual thinking budget; Haiku 5.5 is the first Haiku with effort levels, takes a 1M-token context window (up from 200k) and 128K output (up from 64K), and caches prompts as short as 512 tokens (Haiku 4.5 needed 4,096). The catch: the same text counts as roughly 30% more tokens than on Haiku 4.5.

Three changes that matter

3. Evidence, with sources

Every benchmark number below comes from the table on Anthropic’s launch page, so read it as vendor-reported and directional. The page points to the Haiku 5.5 system card for how the evaluations were run and does not state the harness or tool settings itself; only one entry, Sonnet 5.5 on FrontierCode 1.1, is labelled with its effort level (Xhigh).

The gap to Sonnet 5.5 is the number to remember: 39.2% vs 70.6% on Terminal-Bench 4.0, 72.4% vs 83.9% on OSWorld 2.1, 46.4% vs 52.1% on FrontierCode 1.1. Haiku 5.5 is close on some tasks and far behind on long terminal work. Customer results are quoted by Anthropic and describe their own eval suites.

4. The cost arithmetic

Every figure here is a list price from Anthropic’s pricing page. I am not inventing usage numbers, so the formulas take your token counts. Start with the unit cost of one million input plus one million output tokens and read the ratios.

Output costs 5x input on Haiku 5.5 ($0.50 vs $0.10), and thinking tokens are billed as output tokens. That is why effort, not the price list, is your main cost lever: it changes how many output tokens every call spends.

# 1M input + 1M output tokens (list price)
Haiku 5.5, prompt <= 100k : 0.10 + 0.50  =  $0.60
Haiku 5.5, prompt  > 100k : 0.50 + 2.50  =  $3.00
Haiku 4.5                 : 1.00 + 5.00  =  $6.00
Sonnet 5.5                : 2.00 + 10.00 = $12.00  -> 12.00 / 0.60 = 20x Haiku 5.5
Opus 5.5                  : 4.00 + 20.00 = $24.00  -> 24.00 / 0.60 = 40x Haiku 5.5
(above 100k the gap to Sonnet 5.5 shrinks: 12.00 / 3.00 = 4x)

# one call, prompt <= 100k tokens ("output" includes thinking tokens)
cost = (input x 0.10 + cache_write_5m x 0.125 + cache_read x 0.01
        + output x 0.50) / 1,000,000

# the 100k cliff (cache reads and writes count toward prompt length)
100,000-token prompt : 100,000 x 0.10 / 1M = $0.0100 input
100,001-token prompt : 100,001 x 0.50 / 1M = $0.0500 input  (5x)

# static prefix of P tokens reused N times, 5-minute cache
no cache : N x P x 0.10 / 1M
cached   : (P x 0.125 + (N - 1) x P x 0.01) / 1M
per 1M prefix tokens: N=2 -> $0.200 vs $0.135 | N=10 -> $1.000 vs $0.215

5. Quickstart: effort levels and the worker pattern

Effort is set at output_config.effort, not inside the thinking object, and needs no beta header. Haiku 5.5 defaults to medium, and setting medium explicitly is the same as omitting it. This is Anthropic’s guidance for choosing a level, condensed from its prompting guide:

Below: one worker call on the Messages API with a cached system prompt, then the same worker defined as an Agent SDK subagent under a bigger orchestrator.

import anthropic

client = anthropic.Anthropic()

# Static worker rules. Haiku 5.5 caches prefixes of 512+ tokens.
WORKER_RULES = "...your stable instructions, 512+ tokens..."

response = client.messages.create(
    model="claude-haiku-5-5",
    max_tokens=4096,                  # thinking counts toward max_tokens
    output_config={"effort": "low"},  # the default is "medium"
    system=[
        {
            "type": "text",
            "text": WORKER_RULES,
            "cache_control": {"type": "ephemeral"},
        }
    ],
    messages=[{"role": "user", "content": "Classify this ticket: ..."}],
)

if response.stop_reason == "refusal":
    ...  # a classifier declined; the same request usually refuses again

for block in response.content:  # a thinking block may come first
    if block.type == "text":
        print(block.text)

print(response.usage)  # cache_read_input_tokens, output_tokens, ...

from claude_agent_sdk import query, ClaudeAgentOptions, AgentDefinition

options = ClaudeAgentOptions(
    allowed_tools=["Read", "Grep", "Glob", "Agent"],
    agents={
        "log-triager": AgentDefinition(
            description="Classifies one log file or test report. Use for narrow, high-volume lookups.",
            prompt="Read only the file you are given. Classify each failure and summarize in 5 lines.",
            tools=["Read", "Grep", "Glob"],
            model="claude-haiku-5-5",  # an alias or a full model ID
            effort="low",
        ),
    },
    env={
        "CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH": "1",
        "CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS": "5",
    },
    max_budget_usd=5.0,
)

# async for message in query(prompt="Triage the CI logs in ./logs", options=options):
#     ...

Per Anthropic’s docs, each subagent starts with a fresh context and only its final message returns to the parent, which keeps the orchestrator’s context small. Each subagent also makes its own API requests that count toward total_cost_usd, so cap the fan-out.

6. Honest limits: when not to use Haiku 5.5

Anthropic’s own cost guide is blunt. In its internal, directional measurements, delegating to cheaper workers saved money in only two situations: routine work with a long cost tail, and work larger than one context window. When the job was one dependent chain, or fit in a single context without a long cost tail, the orchestrator’s model alone at lower effort came out ahead in every case measured. Those tests used Sonnet 5 workers, not Haiku 5.5, and the guide says to measure Haiku 5.5 yourself.

7. What to do this week

Do not migrate everything. Pick one narrow, high-volume step of the kind Anthropic names for Haiku 5.5, such as classification, extraction, routing, summarization or quick lookups, and prove it on your own data.

Verdict

Haiku 5.5 is a real worker tier: 20x cheaper than Sonnet 5.5 and 40x cheaper than Opus 5.5 on input and output list prices for prompts under 100k, with effort levels to tune the thinking bill. On Anthropic’s numbers it is far better than Haiku 4.5 and clearly behind Sonnet 5.5 on long terminal work.

Use it for narrow, high-volume steps under an orchestrator, keep prompts under 100k and effort fixed per role, and measure before you widen its job. If the work is one hard chain, pay for the bigger model.