English · 6 min read
Claude Haiku 5.5 vs GPT-6 Luna: both $0.10 / $0.50, so which is the cheapest model?
Claude Haiku 5.5 and GPT-6 Luna list at the same price per token. An independent test found one writes about three times as much to finish a task. Here is how a small business should choose.
By Dapols ·
The short answer: Claude Haiku 5.5 and GPT-6 Luna both list at $0.10 per 1M input tokens and $0.50 per 1M output tokens. For high-volume work where the bill matters most, start with GPT-6 Luna, because an independent test found Haiku 5.5 writes roughly three times as many output tokens per task. Pick Haiku 5.5 when the work is harder and quality matters more than the last cent, and keep its prompts under 100,000 tokens.
Prices are from Anthropic and OpenAI, checked on 20 October 2026. Our arithmetic from those list prices is labelled as such.
The two price cards
| Claude Haiku 5.5 | GPT-6 Luna | |
|---|---|---|
| Input per 1M tokens | $0.10 (prompts up to 100K tokens) | $0.10 |
| Output per 1M tokens | $0.50 (prompts up to 100K tokens) | $0.50 |
| Cached input per 1M | $0.01 (under 100K) | $0.01 |
| Long-prompt rate | $0.50 / $2.50 above 100K tokens | $0.20 / $0.75 for the whole request above 272K tokens |
| Released | 7 Oct 2026 | 22 Sep 2026 |
Haiku 5.5 is the first Claude priced by prompt length, so a prompt over 100,000 tokens costs five times the base input rate. Luna's surcharge starts later, at 272,000 input tokens. For ordinary prompts, a few pages of text, neither threshold matters.
Why the same price is not the same bill
Artificial Analysis, an independent testing firm, scored Haiku 5.5 43 on its Intelligence Index at max effort, ahead of GPT-6 Luna's 37. It also found Haiku 5.5 used about 162,000 output tokens per index task, roughly three times Luna's at about 50,000. Because output is where most of the cost sits, a finished job on Haiku costs more than on Luna at the same per-token price.
Here is what that looks like for a made-up but typical month, using our own arithmetic. Say you send 2 million tokens in:
- If both models wrote 400,000 tokens out: 2 x $0.10 + 0.4 x $0.50 = $0.40 each.
- If Haiku 5.5 wrote three times as much (1.2 million) while Luna wrote 400,000: Haiku comes to 2 x $0.10 + 1.2 x $0.50 = $0.80, double Luna's $0.40.
At these sizes the gap is cents. It matters at scale: a support inbox, a product-description pipeline or a review-sorting job running tens of thousands of times.
What each vendor claims
Anthropic reports 39.2% for Haiku 5.5 on Terminal-Bench 4.0 at max effort, against 16.4% for GPT-6 Luna (VentureBeat). OpenAI reports 66.6% for Luna on the DeepSWE coding benchmark at max effort (OpenAI). Those are different benchmarks, and each is the vendor's own number, so they do not settle which model is better for you.
How to choose
- Sorting, tagging, summarising, short replies, thousands of times: GPT-6 Luna. It is the cheaper finished job, with the lowest list price on our tracker.
- Messier work such as longer reasoning, tool use or agent steps: try Haiku 5.5 on a sample. It scores higher independently, and you can turn its effort setting down to cut the output.
- Very long prompts: neither is cheap. Above 100,000 tokens Haiku jumps to $0.50 / $2.50, and above 272,000 Luna bills the whole request at $0.20 / $0.75. For long documents, DeepSeek V4.1 Flash is cheaper than both at about $0.15 / $0.60 off-peak (see what changed in V4.1 Flash).
- Anything you cannot test first: run 50 real examples through each and compare the actual bill, not the rate card.
See the two side by side on Claude Haiku 5.5 vs GPT-6 Luna, and read our earlier Haiku 5.5 pricing review.
Sources: Anthropic pricing, OpenAI API pricing, Artificial Analysis: Claude Haiku 5.5, Artificial Analysis: GPT-6 Luna, VentureBeat, OpenAI: GPT-6 Sol and Luna, DeepSeek pricing