AI Insights: Latency & Cost Optimization · part 14 of 17
GPT-5.6-Luna: tested and disqualified
Best quality in the whole test — and still no role for it
The question
Does the newly available GPT-5.6-Luna deployment ('1/5th the price') change the model choice?
Best-in-test quality (8.0/7.75), but 1,926 hidden reasoning tokens push TTFT to 18 s and real cost 30% above GPT-5.1.
The pricing illusion
Sticker price is $1/$6 per 1M tokens — nominally a fifth of GPT-5.1's output rate. But Luna is a reasoning model: a median 1,926 hidden reasoning tokens per answer are billed at the $6 output rate and never appear in the response. Those invisible tokens are 71% of the output bill, making the real cost $0.0173 per answer — 30% more than GPT-5.1 — while the user waits 18 seconds for a first word. Streaming can't bring that forward, because reasoning happens before any output token exists.
Results (n=10, V5 prompt)
| Metric | GPT-5.6-Luna | GPT-5.1 + V5 | Qwen W8 |
|---|---|---|---|
| Opus / Sonnet judge | 8.0 / 7.75 | 8.0 / 7.5 | 7.95 / 7.0 |
| TTFT | 18.1 s | ~1.6 s | 614 ms |
| Wall time | 22.2 s | ~15 s | 6.1 s |
| Real cost/answer | $0.0173 | $0.0136 | $0.00176 |