gpt-5.6-luna is the floor price on Claudexia: $0.1 per million tokens, input and output. Three times cheaper than Sonnet, four times cheaper than Opus. There's no Haiku tier here, and Luna fills the same niche: a cheap model for mechanical work.
One thing worth flagging: Luna is a different family, not Claude. The request format is compatible with both Anthropic Messages and OpenAI Chat Completions, but behavior and answer quality on your own prompts should be checked separately — copy-pasting a Sonnet prompt over doesn't guarantee the same result.
Where Luna genuinely earns its keep
Bulk classification. Tag support tickets by category, score review sentiment, sort inbound mail by priority. A task with tight boundaries and a short answer: Luna handles it at a volume where flagship pricing would kill your unit economics.
Extraction. Pull a name, date, amount, and status out of an email or document into JSON. The output shape is fixed ahead of time, the model doesn't need to reason, just find and reformat.
Log parsing. Break an error string into its components, spot a pattern in a stack trace, normalize mismatched log formats into one shape. Repetitive mechanical work at volume.
Format conversion. CSV to JSON, YAML to TOML, one markdown dialect to another. Simple rules, high throughput.
Where Luna falls short
Don't expect Luna to reason at the level of Sonnet or Opus. On multi-step planning, debugging unfamiliar code, or anything that needs several conflicting requirements held in mind at once, it loses the thread noticeably sooner than the bigger models. Don't put Luna at the end of a chain where you need a final synthesis — put it on steps where the answer is short and the rule is clear.
A working example
from openai import OpenAI
client = OpenAI(api_key="sk_cdx_your_key_here", base_url="https://api.claudexia.tech/v1")
resp = client.chat.completions.create(
model="gpt-5.6-luna",
messages=[
{"role": "system", "content": "Classify the ticket into one of: bug, question, feature. Reply with one word."},
{"role": "user", "content": "Save button stopped working after the update."},
],
)
print(resp.choices[0].message.content)
The same call works in Anthropic Messages format too — same key, same endpoint, only the request body shape changes.
What that costs in practice
Take ticket classification at volume. 100,000 tickets a month, a 200-token prompt, a one-word answer at 20 tokens. Input adds up to 20 million tokens, that's $2. Output: 2 million tokens, another $0.2. A little over $2 a month for something a human used to do by hand, or nobody did at all.
The two-stage pipeline
A pattern worth copying: let the flagship define the rule once, then let Luna apply it to a thousand items. Opus or Sonnet reads a sample and writes the classification criteria, then Luna labels the whole batch for pennies. Pay for expensive reasoning once, cheap execution at scale.
Checking it against your own data
Run the same sample through Luna and Sonnet, and check accuracy by hand on a hundred examples. The calculator on the site shows the bill difference at different volumes, and per-key stats tell you exactly how much a given pipeline step is costing.
In short
Luna is cheap by design, built for volume, not for depth. For classification, extraction, parsing, and conversion it's the best money-to-output ratio in the lineup. For reasoning and synthesis, move up a tier, and test behavior separately since it's a different family.