Claude Haiku 4.5

Claude Haiku 4.5 costs $1/$5 per million tokens with a 200k context and 64k output. Where it beats the bigger models on cost per task — and where it quietly fails.

Claude Haiku 4.5 is Anthropic's cheapest and fastest model: $1 per million input tokens, $5 per million output, a 200,000-token context window, 64,000 max output tokens, and the API ID claude-haiku-4-5. Released in October 2025, it is the model you route high-volume classification, extraction and tagging work to — and the one you should stop using the moment a task needs current knowledge or deep reasoning.

This page is cost engineering rather than a spec sheet. It works out what a thousand real calls cost, gives three copy-paste recipes, and is honest about the three ways Haiku 4.5 is now behind the rest of the line-up.

#What do 1,000 classification calls actually cost?

Roughly two dollars at list price, and about 63 cents once you cache the prompt. Here is the arithmetic, using a workload shape that matches most real classification jobs: a shared instruction block and rubric that never changes, a short piece of per-item text, and a one-label answer.

• Shared system prompt and rubric: 1,500 tokens, identical on every call
• Per-item text to classify: 400 tokens
• Output: 15 tokens — a label and nothing else

Two details in that table are easy to miss and both favour Haiku. First, the cache write is a rounding error: 1,500 tokens at 1.25× base input costs about a fifth of a cent, once, while the other 999 calls read the same prefix at $0.10 per million. Caching pays for itself after a single read. Second, Haiku 4.5 still uses the older tokenizer. Sonnet 5 and Opus 5 use the tokenizer introduced with Opus 4.7, which counts roughly 30% more tokens for the same text — which is why the Sonnet and Opus rows show 2.47M input tokens for the same 1,900 tokens of source material per call. The nominal 2× price gap between Haiku and Sonnet is closer to 2.6× in practice.

#Three copy-paste recipes for Haiku 4.5

All three use the current model alias and pass no temperature, top_p or top_k. Install the SDK with pip install anthropic; it reads ANTHROPIC_API_KEY from the environment. More request shapes are on the Claude API page.

#1. Classification with a cached rubric

from anthropic import Anthropic

client = Anthropic()

RUBRIC = open("rubric.md").read()   # must exceed 1,024 tokens to cache

def classify(text: str) -> str:
    msg = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=16,
        system=[{
            "type": "text",
            "text": RUBRIC,
            "cache_control": {"type": "ephemeral"},   # 5-minute TTL
        }],
        messages=[{
            "role": "user",
            "content": f"Classify this ticket. Reply with the label only.\n\n{text}",
        }],
    )
    return msg.content[0].text.strip()

Use "cache_control": {"type": "ephemeral", "ttl": "1h"} if calls are more than five minutes apart. The 1-hour write costs 2× base input and pays for itself after two reads.

#2. Structured extraction with a forced tool call

Asking for JSON in prose gets you JSON most of the time. Forcing a tool call gets you a schema-valid object every time, which matters when you are running a hundred thousand of them.

SCHEMA = {
    "name": "record_invoice",
    "description": "Record the fields extracted from an invoice.",
    "input_schema": {
        "type": "object",
        "properties": {
            "supplier":    {"type": "string"},
            "invoice_no":  {"type": "string"},
            "total_gbp":   {"type": "number"},
            "issued_on":   {"type": "string", "description": "ISO 8601 date"},
        },
        "required": ["supplier", "invoice_no", "total_gbp"],
    },
}

def extract(document: str) -> dict:
    msg = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=512,
        tools=[SCHEMA],
        tool_choice={"type": "tool", "name": "record_invoice"},
        messages=[{"role": "user", "content": document}],
    )
    for block in msg.content:
        if block.type == "tool_use":
            return block.input
    return {}

Note that any tool definition adds system-prompt overhead — 496 input tokens on Haiku 4.5 with tool_choice of auto, 588 when you force a specific tool. At $1 per million that is about six hundredths of a cent per call, but it is worth knowing when you are modelling a million calls.

#3. Routing: let Haiku decide when to spend more

ESCALATE_TO = "claude-sonnet-5"

TRIAGE = (
    "Decide whether this request can be answered from the FAQ alone.\n"
    "Reply with exactly one word: SIMPLE or COMPLEX."
)

def answer(question: str, faq: str) -> tuple[str, str]:
    triage = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=8,
        system=TRIAGE,
        messages=[{"role": "user", "content": question}],
    )
    verdict = triage.content[0].text.strip().upper()

    model = "claude-haiku-4-5" if verdict == "SIMPLE" else ESCALATE_TO
    reply = client.messages.create(
        model=model,
        max_tokens=1024,
        system=faq,
        messages=[{"role": "user", "content": question}],
    )
    return model, reply.content[0].text

The triage call costs a fraction of a cent and is fast enough to be invisible to the user. If 70% of your traffic is answerable by Haiku, you have cut the model line of your bill by roughly two thirds without touching the quality of the hard 30%. The model comparison guide develops this pattern further, including confidence-based escalation and a three-tier ladder, and automation patterns covers running it at scale.

#Where Haiku 4.5 is behind the rest of the line-up

Haiku 4.5 shipped in October 2025 and the models around it have moved on. Three gaps are structural, not cosmetic.

The knowledge cutoff deserves emphasis because it produces confidently wrong answers rather than errors. Haiku 4.5 does not know 2026 happened. It has never heard of Sonnet 5, Opus 5 or Fable 5. Ask it about the current Claude line-up, recent pricing, or anything that changed this year and it will answer from a 2025 world-view without flagging it. Anything time-sensitive must arrive in the prompt — from retrieval, a web search tool, or your own database. The model index lists every cutoff side by side.

Two smaller limitations round it out. The computer use and browser use tools that went GA on 19 August 2026 are available on Fable 5, Opus 5, Sonnet 5 and Opus 4.8 — not on Haiku 4.5. And the code execution tool runs on Haiku 4.5, but the newer versions behave like the original: no REPL state persistence, no programmatic tool calling. If your agent depends on either, Haiku cannot be the model driving it, though it can still be a subagent underneath one. See MCP and Claude agents for how that decomposition works.

#When not to use Haiku 4.5

The honest list. Each of these is a case where the cheap model costs more, because retries, review time or a wrong answer in production dwarf the token difference.

• Anything that depends on 2026 facts. Covered above, and the most common way people get burned.
• Long documents. 200k tokens is a lot of text but it is a fifth of what the current tier above holds. Contract sets, whole codebases and research corpora that fit comfortably in Sonnet 5 will not fit here.
• Multi-step agentic work. Long-horizon tool loops are what adaptive and interleaved thinking exist for, and Haiku 4.5 has neither. Anthropic positions it for real-time and sub-agent tasks, not for driving the loop.
• Work where being wrong is expensive. Medical, legal, financial or safety-relevant output. The saving is measured in cents; the exposure is not.
• Nuanced writing. For customer-facing prose, tone and judgement matter more than latency. See Claude as a writing assistant for where the larger models earn their price.
• Genuinely hard reasoning. If a task needs deep, deliberate reasoning, buy it: adaptive thinking on a bigger model will usually beat a longer thinking budget on a smaller one.

For historical context, Anthropic reported Haiku 4.5 at 73.3% on SWE-bench Verified at launch in October 2025, averaged over 50 trials with a 128k thinking budget — a figure that placed it above Opus 4.1 from two months earlier. That benchmark is no longer reported by frontier vendors, so treat it as a snapshot of October 2025 rather than a current ranking. Haiku's own version history is archived on the Haiku page, and how per-token rates relate to subscription plans is covered under pricing.

#Frequently asked questions