GPT-5.6 Luna vs Gemini 3.6 Flash: The Cheaper Model Has the Higher Price

The two obvious budget-tier choices right now, by list price:

Luna is cheaper on both dimensions — 33% less on input, 20% less on output. On a rate card it's not close.

We ran them head to head in a single session, same prompts, same max_tokens, same grading code.

Both handled everything

Both budget models found a bug that stays invisible on ordinary input, counted characters correctly, and held two simultaneous output constraints. This is consistent with everything else in this series: on verifiable one-shot work, the cheap tier is not a downgrade.

The cheaper model cost more

Gemini 3.6 Flash cost 32% less — while charging more per token on both input and output.

Luna used 1.9× the output tokens for identically graded answers, and that swamps a 20% rate advantage. This is the fifth time in this series the rate card has pointed the wrong way, and it's the cleanest case yet: the model that is more expensive on every line of the price list produced the cheaper invoice.

If you picked Luna off a pricing page to save money on a high-volume pipeline, you are paying about 47% more than the alternative you skipped.

Luna is twice as fast

Same mechanism, opposite direction: Gemini 3.6 Flash is a notably slow Flash model — slower than the 3.5 Flash it replaced — while Luna returns quickly despite being wordier.

So the budget tier splits cleanly, and not along the axis the names suggest:

• Cheapest per unit of work: Gemini 3.6 Flash
• Fastest: GPT-5.6 Luna, by 2×

Constraint adherence goes to Gemini, decisively

Ten runs each of the no-letter-"e" test:

Ten for ten against seven for ten. At this sample size that gap is larger than the run-to-run noise we've documented elsewhere, and it matches Gemini 3.6 Flash's perfect record in our previous test — twenty clean runs across two sessions.

For reference, Claude Opus 5 sits at 13 of 29 on this same prompt, at roughly ten times the price.

How to choose

Default to Gemini 3.6 Flash. Cheaper in practice, and the most reliable constraint-follower we've measured at any price. For batch processing, classification, extraction, and anything with a hard output rule, it wins on the two axes that matter.

Choose Luna when latency is the product. Twice as fast is a real difference in an interactive path, and 47% more on a small number is often an easy trade. Just make that choice deliberately rather than because the price list looked lower.

If you want both, look one tier over. Grok 4.5 came in at $0.003272 on this same battery — level with Gemini 3.6 Flash on cost — and was faster. Its list rate ($2/$6) suggests it shouldn't compete with the budget tier on price. It does, because it's exceptionally terse.

And measure your own prompts. Every surprise in this post came from token counts, not rates. Our tasks produce short answers; if yours produce long ones, the verbosity gap grows and the ranking may shift further.

Run it yourself

No account, no API key:

• The constraint test — Gemini's clean sweep, Luna's 7/10
• The letter-counting task — answer is 10

To compare both on your own prompts with real token counts and latency, that's Promptster.

Run 2026-08-09 against gpt-5.6-luna and gemini-3.6-flash via the Promptster API, max_tokens: 4000. Both models measured in the same session, so the latency comparison shares network conditions — unlike figures quoted across posts, which are not controlled for that. Costs computed from list pricing: Luna $1/$6, Gemini 3.6 Flash $1.50/$7.50 per million input/output tokens. The three graded tasks were run once per model; the constraint task ten times per model.