Anthropic released Claude Haiku 5.5 on 7 October 2026, completing the Claude 5.5 family alongside Opus 5.5 and Sonnet 5.5. Haiku is the small, fast model, built for high-volume work such as classification, extraction, routing, summaries and "subagent" tasks inside bigger AI workflows. The headline is price: on prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, a tenth of Haiku 4.5's list price.
Pricing in detail
Claude Haiku 5.5 prices per million tokens
- Input, prompts up to 100K
- $0.10
- Output, prompts up to 100K
- $0.50
- Input, prompts over 100K
- $0.50
- Output, prompts over 100K
- $2.50
- Cache reads
- $0.01
- Cache writes (5 minutes)
- $0.125
- Batch processing
- 50% off
- Haiku 4.5, for comparison
- $1 input and $5 output
Anthropic says about 90% of Haiku 4.5 requests stayed under 100,000 tokens, so most workloads get the low price. The minimum prompt you can cache also drops from 4,096 tokens to 512, which makes caching worthwhile for short, repeated system prompts too.
A worked example: one million support tickets
Say you classify one million tickets, each 1,000 input tokens on Haiku 4.5's tokenizer and a short 50-token label. Haiku 5.5's tokenizer counts the same text as roughly 30% more tokens, so it is 1,300 input and 65 output tokens per ticket.
Haiku 4.5Haiku 5.5
List prices, before any thinking tokens, caching or batch discount. With batch processing, Haiku 5.5 would be about $81.
Benchmarks
Anthropic's own results show a huge jump over Haiku 4.5 on agentic work, where a model has to use a computer or a terminal to finish a task, and a clear lead over OpenAI's small model, GPT-6 Luna.
Higher is better. Anthropic-reported. Sonnet 5.5 scores higher than all three.
Higher is better. Anthropic-reported.
On Humanity's Last Exam, a very hard reasoning test, it scores 45.9% without tools and 57.4% with tools. On the independent Artificial Analysis Intelligence Index it scores 43, at about $0.21 per benchmark task, which makes it one of the cheapest capable models available. For context, see the best AI models in October 2026.
Haiku 5.5 versus its closest rivals
Claude Haiku 5.5
Anthropic- Price
- $0.10 / $0.50 up to 100K
- Context
- 1M tokens
- Output
- 128K tokens
- Inputs
- text and images
- Strongest of the three on agentic benchmarks
- Five times the price above 100K tokens
GPT-6 Luna
OpenAI- Price
- $0.10 / $0.50 (higher above 272K)
- Context
- 1.05M tokens
- Output
- 128K tokens
- Inputs
- text and images
- Cheaper for very long prompts
- Lower agentic scores in Anthropic's tests
Gemini 3.5 Flash-Lite
Google- Price
- $0.30 / $2.50, flat
- Context
- about 1M tokens
- Output
- 65K tokens
- Inputs
- text, image, video, audio and PDF
- Widest range of inputs
- Higher list price
Where it fits for data teams
- Subagents
- A big model plans; Haiku does the many small lookups. Rogo's example: Haiku pulls a revenue line from a 10-K while a larger model builds the deck.
- Classification and tagging
- Label tickets, reviews or transactions at scale, ideally with batch processing for half price.
- Extraction
- Turn invoices, contracts or emails into structured rows.
- Summaries of long documents
- 1M tokens means whole reports or long threads in one prompt.
- Fast assistants
- Live support and browser tasks where speed matters more than depth.
For complex agentic coding or deep analysis, Anthropic itself recommends Sonnet 5.5 or Opus 5.5.
Using it, and what breaks when you upgrade
The model ID is claude-haiku-5-5, available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, and in preview in Snowflake Cortex.
import anthropic
client = anthropic.Anthropic()
msg = client.messages.create(
model="claude-haiku-5-5",
max_tokens=2000, # leave room for thinking tokens
messages=[{"role": "user",
"content": "Classify this ticket as billing, bug or other: ..."}],
)
for block in msg.content: # responses can start with a thinking block
if block.type == "text":
print(block.text)Breaking changes from Haiku 4.5
- temperature, top_p, top_k
- Non-default values return an error. Remove them.
- Prefill
- Ending messages with an assistant turn returns an error. End with a user turn.
- Manual thinking budgets
- budget_tokens returns an error. Use adaptive thinking and effort.
- Tokenizer
- About 30% more tokens for the same text. Recount prompts and budgets.
- Response shape
- Select content blocks by type, not position.
- Old computer-use tool
- computer_20250124 returns an error. Use the newer tool version.
Check your understanding
1/3What does Claude Haiku 5.5 cost for input on a 50,000-token prompt?
The $0.50 input price applies only above 100K tokens.
2/3Why can the same prompt cost more tokens than on Haiku 4.5?
Recount prompts and revisit max_tokens when migrating.
3/3What happens if you send temperature=0.2?
Non-default sampling parameters are no longer supported.
You got 0 out of 3.
Comparing models for a project? Read the best AI models in October 2026, GPT-6.1 Sol and DevDay 2026 and Gemini 4 Argon explained.
Sources: Anthropic's "What's new in Claude Haiku 5.5" documentation and system card figures as reported by MarkTechPost (7 October 2026), and the Artificial Analysis leaderboard (mid-October 2026). Benchmarks are vendor-reported unless stated.
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 17 October 2026. About me