Anthropic released Claude Haiku 5.5 on 7 October 2026 at $0.10 per million input tokens and $0.50 per million output tokens, exactly what OpenAI charges for GPT-6 Luna. In Anthropic’s own table, Haiku 5.5 scores higher than Luna on every benchmark the two share. For short prompts it is 90% cheaper than Claude Haiku 4.5, which puts Anthropic back in the running for high-volume work it used to lose on price.
Same price as Luna, much higher scores on desktop and terminal work
Haiku is Anthropic’s smallest model line, built for jobs you run thousands of times a day. Version 5.5 reads text and images, has a context window of 1 million tokens and writes up to 128,000 tokens per answer. It is also the first Haiku with an effort setting, a dial for how long the model thinks before it answers. The default is medium.
Anthropic’s table shows the widest gaps on tasks where the model has to operate a computer or a terminal, and the smallest gap on pure coding.
| What is measured | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| Input price per million tokens | $0.10 (prompts up to 100K) | $0.10 (prompts up to 272K) |
| Output price per million tokens | $0.50 | $0.50 |
| OSWorld 2.1, offline subset (operating a desktop) | 72.4% | 48.9% |
| Terminal-Bench 4.0 (command-line tasks) | 39.2% (Haiku 4.5: 0.0%) | 16.4% |
| FrontierCode 1.1 Main (coding) | 46.4% | 42.4% |
| GDPval-AA v2.1 (office work, rating) | 1620 | 1437 |
All scores come from Anthropic’s announcement. Anthropic refers to the system card for its methods but does not say whether it ran GPT-6 Luna itself or took OpenAI’s figures. The Luna prices are OpenAI’s list prices as reported by MarkTechPost. Claude Sonnet 5.5 still leads every row, with 70.6% on Terminal-Bench 4.0, and Anthropic itself points complex coding agents to Sonnet 5.5 and Opus 5.5.
Prompts over 100,000 tokens cost five times more
The headline price only holds for prompts up to 100,000 tokens. Above that, Haiku 5.5 charges $0.50 for input and $2.50 for output per million tokens. GPT-6 Luna keeps its $0.10 input rate up to 272,000 tokens and then moves to $0.20. If you feed whole contracts or long transcripts into a small model, Luna stays far cheaper on paper.
The second catch is the tokenizer, the part that cuts text into billable pieces. Haiku 5.5 uses Anthropic’s newer one, which counts about 30% more tokens for the same text than Haiku 4.5. Anthropic puts the average saving over Haiku 4.5 at about 75% rather than 90% for that reason. It adds that about 90% of Haiku 4.5 requests stayed under 100,000 tokens, so most current users land in the cheap bracket.
Where to run it, according to Anthropic’s documentation:
- Model ID: claude-haiku-5-5 (on Amazon Bedrock: anthropic.claude-haiku-5-5)
- Platforms: Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS
- Cache reads: $0.01 per million tokens; batch jobs at half price
- Knowledge cutoff: June 2026
- Leave temperature, top_p and top_k at their defaults, or the API returns an error
- Retirement no earlier than 7 October 2027
Anthropic also halved the cache-read price of Claude Sonnet 5.5 to $0.10 per million tokens, which it says cuts the cost of most agent tasks by about 20%. The announcement says nothing about EU data residency. If your data has to stay in Europe, check whether your cloud provider offers Haiku 5.5 in an EU region before you plan around it.
Move short, high-volume Claude calls to Haiku 5.5 now, and price long prompts first
This release matters most to teams that already run Claude and pay Haiku 4.5 or Sonnet prices for routine steps: ticket triage, field extraction, summaries, and the small helper agents a larger model sends off to look things up. For those calls the switch is a large saving, as long as quality holds. Reports from Asana, HubSpot, Box and AlphaSense all point that way, but Anthropic chose them, so they show what is possible rather than what is typical.
Test one real workload before you switch it over. Count tokens on your own text with the new tokenizer, run a few hundred cases next to your current model, and compare accuracy and cost per task. If your prompts often pass 100,000 tokens, run the same test on GPT-6 Luna, because there the price advantage flips. For organisations building agent systems, the setup to try is a strong model that plans and Haiku 5.5 doing the many small steps. Cognition and Rogo describe exactly that pattern.
For more information, visit the official announcement of Claude Haiku 5.5 on the Anthropic website.
