Claude Haiku 5.5 vs GPT-6 Luna: API pricing, caching, and long prompts
For the half cached 100,000 input, 5,000 output request, both models cost $0.008. At volume, the long prompt rules and your cache hit rate move the total as much as the base rates do.
Small models get bought by the million requests, so a small gap in a rate turns into a real line on the invoice. Claude Haiku 5.5 charges $0.10 per million input tokens and $0.50 for output, GPT-6 Luna charges $0.10 and $0.50, and neither is lower on both, so your input to output ratio decides. The rate card isn't the whole story, though. Haiku moves a request to a higher tier once the prompt passes 100,000 tokens, and Haiku counts cache reads toward that line. Luna has a long prompt tier of its own above 272,000 tokens. Those rules matter most for agents and document pipelines, which is where small models earn their keep.
Small fast models at volume, with cache reads, batch jobs, and Haiku's long prompt tier in the math.
By TechCompare · Updated
Anthropic: Claude Haiku 5.5
Anthropic's own price
- Input
- $0.10
- Output
- $0.50
- Cached
- $0.01
per 1M tokens
Price checked 2026-11-07
OpenAI: GPT-6 Luna
OpenAI's own price
- Input
- $0.10
- Output
- $0.50
- Cached
- $0.01
per 1M tokens
Price checked 2026-11-07
Cost Comparison
Based on 100,000 input tokens (50% cached), 5,000 output tokens, and 100 requests.Prices come live from OpenRouter at each maker's own price where it's listed. Temporary deals are left out.
Side-by-side specs
| Spec | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| Standard input per 1M tokens | $0.10 | $0.10 |
| Standard output per 1M tokens | $0.50 | $0.50 |
| Cache reads per 1M tokens | $0.01 | $0.01 |
| 100K input + 5K output, no cache hits | $0.0125 | $0.0125 |
| 100K input + 5K output, half the input cached | $0.008 | $0.008 |
| Moderation, 10M requests of 1.2K input, 50 output | $1,450 | $1,450 |
| Agent fleet, 200K calls of 70K input, 1.5K output | $416 | $416 |
| Input / output on long prompts, per 1M | $0.50 / $2.50 above 100,000 input | $0.20 / $0.75 above 272,000 input |
| Batch API input / output per 1M | $0.05 / $0.25 | $0.05 / $0.25 |
| Context window | 1,000,000 tokens | 1,050,000 tokens |
| Input types | Text, image, file | Text, image, file |
How they differ
Four workloads at volume show where the rates land. A moderation or classification endpoint that handles 10,000,000 requests a month, each with 1,200 input tokens and 50 output tokens and no cache hits, costs $1,450 on Haiku 5.5 and $1,450 on Luna. A support chat product with 1,000,000 conversations a month, each with 5,000 input tokens and 400 output tokens and half the input cached, costs $475 and $475. A browser use or subagent fleet that runs 5,000 sessions of 40 calls, each call carrying 70,000 input tokens and 1,500 output tokens with nine in ten input tokens read from cache, makes 200,000 calls and costs $416 and $416. A document pipeline that sends 300,000 input tokens and 3,000 output tokens per request with no cache hits costs $0.1575 and $0.06225 per request. Haiku bills the whole request at $0.50 input, $0.05 for cache reads and $2.50 output once a prompt passes 100,000 tokens. Prompt length includes cache reads and writes, so a context that's mostly cached still counts toward that line. Compare 100,000 with the 70,000 input tokens each agent call carries here, because the Haiku total uses whichever tier applies to a prompt that size. Luna moves the whole request to $0.20 input and $0.75 output above 272,000 tokens. Both list batch prices, which bring the 10,000,000 request classification job to $725 on Haiku and $725 on Luna. Haiku's five minute cache writes cost $0.125 per million tokens, and one hour writes cost $0.20. This page doesn't price Luna's cache writes, so read OpenAI's current terms before you design a cache heavy system around it. The examples count cache reads only.
Verdict
Price your real traffic, not the headline rates. For short, uncached calls such as classification, the input and output rates are nearly everything. For agents, the cache read rate and the prompt length rules take over. Both read cached input at $0.01 per million tokens, so the cache rate won't separate them for an agent that resends a big prefix on every call. Then run both on a sample of your own tasks. A small model that needs a retry on one call in five has a different bill from its rate card, and the one that's wrong less often on your data can win even when its rates aren't the lower ones. Count billed tokens per finished task, not per call, and keep a note of where your longest prompts sit relative to Haiku's tier line.
Which should you pick?
Choose Claude Haiku 5.5
Choose Haiku 5.5 when you run high volume, cost sensitive work such as summarization, subagents and browser use, and you can keep most prompts under its long prompt line. It takes text, image and file input, has a 1,000,000 token context window and allows up to 128,000 output tokens. Batch prices are listed today, which suit overnight jobs. Agents that hold a long, stable context should model the tier rule carefully, since cache reads and writes count toward prompt length. Check Anthropic's model overview for current limits before you commit.
Claude Haiku 5.5 model overviewChoose GPT-6 Luna
Choose Luna when your work is chat, classification or light agentic tasks and you're already building on OpenAI's APIs. It takes text, image and file input and has a 1,050,000 token context window. It has a long prompt tier above 272,000 tokens, so price big requests at that rate. Batch prices are listed today. Open the Luna pricing page on this site for its own worked examples, and test on your real prompts, because cost per finished task is what you'll be billed for.
GPT-6 Luna pricing