Claude Haiku 5.5 Launches: 1M Context, 90% Cheaper

​

Key Takeaways

• About one-tenth the price: for prompts up to 100K tokens, input is $0.10 and output $0.50 per million tokens, with cache reads at $0.01. Anthropic estimates these requests cost about 90％ less than on Haiku 4.5, even after counting the extra tokens from the new tokenizer
• A generational capability jump: Anthropic reports GDPval-AA knowledge work at 1620 (Haiku 4.5: 735), OSWorld 2.1 computer use at 72.4% (Haiku 4.5: 15.7%), and Terminal-Bench 4.0 at 39.2% (Haiku 4.5: 0%)
• 1M context, 128K output: up from 200K context and 64K output on Haiku 4.5. It is also the first Haiku with effort levels (low through max)
• 5× unit price above 100K tokens: input $0.50 / output $2.50. Run the numbers before using it for long documents or long conversations
• Live on APIYI: claude-haiku-5-5 and claude-haiku-5-5-thinking, on both the OpenAI-compatible and native Anthropic endpoints, with both pricing tiers matching the official list item by item

​

Background

​

Detailed Analysis

​

Key Features

​

Benchmarks (vendor-reported)

• Against its price peer: Haiku 5.5 beats OpenAI’s small model GPT-6 Luna on every benchmark Anthropic lists, and more than doubles it on Terminal-Bench 4.0
• Against Sonnet 5.5: computer use (OSWorld 72.4% vs 83.9%) reaches about 85% of Sonnet 5.5, and knowledge work trails by about 220 Elo; agentic coding (Terminal-Bench 39.2% vs 70.6%) is where the gap is widest
• Customer reports: Box saw scores 11 points above Haiku 4.5 at about half the latency; Asana reported more than 30% lower task-completion latency

​

Independent evaluation (Artificial Analysis)

​

Developer notes: breaking changes from Haiku 4.5

• Responses can start with a thinking block: adaptive thinking is on by default, so a thinking block can appear even if the request never mentions thinking. Select text blocks by type instead of reading the first block
• Thinking tokens count toward max_tokens: a small max_tokens carried over from Haiku 4.5 can stop right after thinking, before any text. Raise it or lower the effort
• Thinking text is omitted by default: set thinking: {"type": "adaptive", "display": "summarized"} if you want a summary
• Forced tool use works: tool_choice set to any or a named tool does not error (it does on Sonnet 5.5), but the model skips thinking and calls the tool directly
• Safety classifiers: refusals come back as stop_reason: "refusal" with no server-side fallback to another model, so handle it in your client

​

Specifications

​

APIYI test results

​

Practical Use

​

Recommended scenarios

1. Classification, extraction, routing: the primary use case. High-volume, short requests under 100K tokens land squarely in the cheapest tier
2. Subagent for larger models: handle lookups, summaries, and data pulls inside coding or research workflows led by Opus 5.5 or Sonnet 5.5
3. Compaction and summaries: long-conversation compaction, meeting notes, ticket summaries, fast and cheap
4. Replacing Haiku 4.5: most Haiku 4.5 workloads get noticeably stronger and cheaper, but check the migration items below first

​

Code examples

​

Native Anthropic format

import anthropic

client = anthropic.Anthropic(
    api_key="sk-your-apiyi-key",
    base_url="https://api.apiyi.com"
)

message = client.messages.create(
    model="claude-haiku-5-5",
    max_tokens=4000,                   # thinking tokens count here too, don't set it too low
    output_config={"effort": "low"},   # default is medium; low is enough for classification
    messages=[
        {"role": "user", "content": "Classify this ticket as: billing / technical issue / account / other. Output the category only.\n\nTicket: My top-up didn't show up in my balance"}
    ]
)

for block in message.content:          # pick text by type, don't read content[0] directly
    if block.type == "text":
        print(block.text)

​

OpenAI-compatible format

from openai import OpenAI

client = OpenAI(
    api_key="sk-your-apiyi-key",
    base_url="https://api.apiyi.com/v1"
)

response = client.chat.completions.create(
    model="claude-haiku-5-5",
    # don't send temperature / top_p / top_k
    messages=[
        {"role": "user", "content": "Extract the company name, amount, and date from the text below as JSON."}
    ]
)

print(response.choices[0].message.content)

​

Six-step checklist for migrating from Haiku 4.5

1. Change the model ID: claude-haiku-4-5-20251001 → claude-haiku-5-5 (no date suffix)
2. Recount tokens: the new tokenizer yields about 30% more tokens for the same text, so recompute max_tokens and cost estimates
3. Replace the thinking config: swap budget_tokens for adaptive thinking plus effort
4. Drop sampling parameters and prefill: don’t send temperature / top_p / top_k, and end messages with a user turn
5. Select content blocks by type: responses can begin with a thinking block
6. Handle refusals: detect stop_reason: "refusal" and fall back to another model yourself if needed

​

Pricing and Availability

​

Pricing

​

Groups and endpoints

​

Stack the top-up bonus

​

Summary and Recommendations

1. On Haiku 4.5 today: upgrade, but run the six-step checklist first; sampling parameters, prefill, and budget_tokens all return a 400
2. Running narrow tasks on Sonnet or Opus: hand lookups, summaries, and classification off to Haiku 5.5 and keep the larger model for decisions; overall cost drops noticeably
3. Keep costs in check: start effort at low or medium, and keep single inputs under 100K tokens to avoid the 5× tier