What it is
Claude Haiku 5.5 is Anthropic's smallest current model, released on 7 October 2026. It is built for high-volume jobs such as classification, data extraction, routing, customer support and acting as a helper (subagent) under a bigger model. It reads text and images, writes text, and has a 1M-token context with 128K output and a knowledge cutoff of June 2026.
Pricing depends on prompt length: $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, and $0.50/$2.50 above that. Cache reads cost $0.01 per million and the Batch API halves both rates. It is the first Haiku with an adjustable effort setting, from low to max.
It is available through the Claude API, the Claude apps, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry. Anthropic commits to keeping it available until at least 7 October 2027.
Our review
Haiku 5.5 has made the cheap tier competitive with mid-priced models. On Artificial Analysis's Intelligence Index it scores 43 at max effort, ahead of Gemini 3.8 Flash at high reasoning (41) and GPT-6 Luna at max (38), and it costs $0.21 per index task against Gemini 3.8 Flash's $1.24. It is also the fastest of the three on that board, at 237 output tokens a second.
Anthropic's launch table makes the jump from Haiku 4.5 plain: 39.2% on Terminal-Bench 4.0 against 0%, and 72.4% on an offline OSWorld 2.1 subset against 15.7%. It beat GPT-6 Luna on every benchmark where both were listed. Those are vendor-run figures, but the independent index points the same way.
The headline price needs two corrections. A newer tokenizer counts the same text as about 30% more tokens than Haiku 4.5 did, which Anthropic says leaves the real saving near 75% rather than 90%. And the rate quintuples once a prompt passes 100,000 tokens, so reading a 300,000-token document costs $0.15 in input where a 90,000-token one costs under a cent.
Effort is where buyers will spend or save. At low effort the index score falls to 29 and the cost per task to $0.02; at max it is 43 and $0.21. Sonnet 5.5 still leads at the top end (56 at max) and on Terminal-Bench (70.6%), which is why Anthropic itself pitches Haiku as a subagent under Sonnet or Opus rather than a replacement for them.
Where it earns its place
A 40-agent support team handling 1M customer replies a month
Draft replies and route tickets in real time
Run Haiku 5.5 at medium effort with a cached system prompt. At about 2,000 input and 300 output tokens per reply that is roughly $350 a month at list price; the same workload on Sonnet 5.5's $2/$10 is about $7,000.
A platform engineer running a coding agent on Opus or Sonnet 5.5
Offload file search, summaries and lookups to a cheaper subagent
Route subagent calls to claude-haiku-5-5; Cognition told Anthropic that Devin Fusion keeps a 66.2 FrontierCode score with Haiku as the sidekick at lower cost and latency. Keep subagent prompts under 100K tokens to stay on the cheaper rate.
A fintech data team extracting figures from 50,000 10-K filings
Pull numbers into a database overnight
Use the Batch API at half price and chunk filings to stay under the 100K-token threshold; Rogo uses Haiku for this kind of 10-K lookup, per Anthropic. Whole filings sent in one prompt move to $0.50/$2.50 before the batch discount.
Before you commit
- Setup
- Change the model ID to claude-haiku-5-5, drop temperature, top_p and top_k (the docs say to omit them), and set effort, which defaults to medium.
- Integrations
- Claude API and SDKs, Message Batches API and prompt caching; also Amazon Bedrock (anthropic.claude-haiku-5-5), Claude Platform on AWS, Google Cloud and Microsoft Foundry.
- Data and privacy
- API inputs and outputs are not used for training and are deleted within 30 days, unless flagged for a Usage Policy breach (kept up to 2 years) or covered by a separate zero-data-retention agreement.
- Cost at scale
- 1M short support replies a month (2,000 tokens in, 300 out) is about $350 at list price or $175 on Batch. A 10,000-token system prompt served from cache costs about $100 across 1M calls instead of $1,000 uncached.
- Limits
- 1M-token context and 128K output (300K on Batch with a beta header). The Start tier allows 1,000 requests and 2M uncached input tokens a minute, with a $500 monthly spend cap that pauses the API until the next month; retirement not before 7 October 2027.
What's good
- $0.10/$0.50 per million tokens under 100K, cache reads $0.01
- 43 on Artificial Analysis's index at max effort, above Gemini 3.8 Flash
- 237 output tokens a second at max effort
- Five effort levels from low to max
- 1M-token context, 128K output
What's not
- Price rises 5x for prompts over 100K tokens
- New tokenizer counts about 30% more tokens
- Well behind Sonnet 5.5 on Terminal-Bench (39.2% vs 70.6%)
- Text and images in only; no audio or video
Best for
High-volume classification, extraction, support and subagent work on short prompts.
Alternatives
Claude Haiku 5.5: common questions
- Is Claude Haiku 5.5 worth it?
- Haiku 5.5 outscores Gemini 3.8 Flash on independent tests at a sixth of the cost per task, which resets what anyone should expect from a budget model. We rate it 4.5 out of 5.
- How much does Claude Haiku 5.5 cost?
- API: $0.10 input / $0.50 output per 1M tokens for prompts up to 100K tokens; $0.50 / $2.50 above that; cache reads $0.01; Batch API 50% off.
- What are the downsides of Claude Haiku 5.5?
- Price rises 5x for prompts over 100K tokens; new tokenizer counts about 30% more tokens; well behind Sonnet 5.5 on Terminal-Bench (39.2% vs 70.6%); text and images in only; no audio or video.
- Who is Claude Haiku 5.5 best for?
- High-volume classification, extraction, support and subagent work on short prompts.
- How much does Claude Haiku 5.5 cost for a team or in production?
- 1M short support replies a month (2,000 tokens in, 300 out) is about $350 at list price or $175 on Batch. A 10,000-token system prompt served from cache costs about $100 across 1M calls instead of $1,000 uncached.
- Is Claude Haiku 5.5 safe for company data?
- API inputs and outputs are not used for training and are deleted within 30 days, unless flagged for a Usage Policy breach (kept up to 2 years) or covered by a separate zero-data-retention agreement.
- What are the best Claude Haiku 5.5 alternatives?
- The closest alternatives we have reviewed are GPT-6 Luna, Gemini 3.5 Flash-Lite and Claude Sonnet 5.5.
Answers checked Oct 22, 2026.
Make Claude Haiku 5.5? Show this rating on your site
