Claude Haiku 5.5

TL;DR Highlight

Haiku 5.5 costs about 75% less than Haiku 4.5 on average and runs faster.

Who Should Read

A backend developer using Haiku 4.5 or GPT-6 Luna for summarization, classification, or subagents who wants to cut cost and latency.

Core Mechanics

Average cost is about 75% below Haiku 4.5
- Requests up to 100k tokens are priced 90% lower than Haiku 4.5.
- Requests over 100k tokens are priced 50% lower.
- The 75% average also accounts for the new tokenizer using slightly more tokens per task.
Pricing splits at 100k tokens
- Up to 100k: $0.10 input and $0.50 output per million tokens.
- Over 100k: $0.50 input and $2.50 output.
- 90% of Haiku 4.5 requests were under 100k.
Haiku now has an effort setting for the first time
- Five levels from Low to Max let you trade cost against intelligence.
- Anthropic published accuracy-vs-cost charts per level for OSWorld, GDPval-AA, and Humanity's Last Exam.
Benchmarks jumped over Haiku 4.5
- OSWorld 2.1 (offline subset) is 72.4% vs 15.7% for Haiku 4.5.
- Terminal-Bench 4.0 is 39.2% vs 0.0% for Haiku 4.5, and below Sonnet 5.5's 70.6%.

Community Reactions

A direct cost comparison with Luna came out higher for Haiku
- One commenter's classification run used 11,893,643 input tokens on Haiku 5.5 vs 7,903,468 on Luna.
- Batch cost was $0.21 for Haiku 5.5 and $0.16 for Luna.
Plotly's data analysis benchmark results
- A commenter reported it was 9x cheaper than Haiku 4.5 and 2 letter grades better.
- It got the straightforward questions right but fell short on deeper statistical analysis.
- GPT-6 Luna did slightly better at about 30% of Haiku 5.5's cost.
Poor results on latency-sensitive internal tasks
- A commenter testing 5.5 as a replacement for Haiku 4.5 reported bad results, including prompt leakage.
- They also found it slower.
Strong value in a large RAG system
- A commenter running a RAG system over 3000+ Cloudflare docs said it had the best quality*speed/price ratio of 100+ models tested.

How to Apply

A/B test your current summarization and classification calls
- Pull 20 to 50 real prompts today and run them on your current model and claude-haiku-5-5 under the same conditions.
- Record accuracy, latency, and input/output token counts together.
- Compare total cost per task rather than unit price, since the new tokenizer can increase token counts.
Raise effort from Low to find the minimum level
- Start each task at Low and lock in the lowest level that clears your accuracy bar.
- Measure cost and latency at each level too.
Keep prompts under 100k tokens
- Check your request logs for the share of prompts over 100k tokens.
- Requests above that pay 5x rates, so shrink them with chunking or summarization.
Split out subagents and compaction to Haiku
- Keep the main agent on Sonnet or Opus and move only summarization, lookup, and compaction steps to Haiku 5.5.
- After the switch, verify that end-to-end task success does not drop.

Try It

copy & run

# Paste into Claude or ChatGPT: check model placement and cost before adopting Haiku 5.5

I run a service that uses LLM APIs. Review the call list below and decide which model tier (Haiku-class / Sonnet-class / Opus-class) fits each call.

[Call list]
{{call name, avg input tokens, avg output tokens, calls per day, task type}}

[Pricing (per 1M tokens)]
Haiku 5.5: $0.10 input / $0.50 output (prompts up to 100k), $0.50 input / $2.50 output (over 100k)
Haiku 4.5: $1.00 input / $5.00 output

[Rules]
- Narrow tasks like summarization, classification, DB queries, compaction, and subagents are Haiku candidates.
- Complex multi-step agentic coding is a Sonnet/Opus candidate.
- Account for Haiku 5.5's new tokenizer using slightly more tokens per task.

[Request]
1. Recommended tier per call with a one-line reason
2. Monthly cost change when moving from Haiku 4.5 to 5.5
3. The 3 calls to A/B test first and the metrics to track

Next step: run the 3 recommended calls on claude-haiku-5-5 at Low, Medium, and High effort and compare accuracy and cost in a table.

Terminology

Effort settingA dial for how long and deeply the model thinks. Turning it up raises accuracy, cost, and time.

Cache readThe cheap rate you pay when re-reading a stored prompt prefix. It makes up a big share of cost in agents that resend the same context.

CompactionShrinking a long conversation or work log into a short summary to cut context length. Like carrying the abridged edition instead of the whole book.

SubagentA helper agent that handles small tasks delegated by the main agent. Like a lead handing pieces of work to a junior.

TokenizerThe rule that splits text into tokens the model reads. Changing the rule changes token counts for the same text, and therefore cost.