GPT-3.5 Turbo vs GPT-5.6 Luna
Where to go when the cheapest old model is switched off.
Side by side
| Attribute | GPT-3.5 Turbo | GPT-5.6 Luna |
|---|---|---|
| Input, per 1M tokens | $0.50 | $0.20 |
| Output, per 1M tokens | $1.50 | $1.20 |
| Context window | 16,385 | 1,050,000 |
| Provider | OpenAI | OpenAI |
| Tokenizer | cl100k_base | o200k_base |
| Long-context tier | none | above 272,000 |
| Retires | 2026-10-23 | no date announced |
Which to choose
- On cost, GPT-5.6 Luna is 2.5x cheaper per million input tokens. On a workload that reads far more than it writes, that ratio is close to the whole difference in the bill.
- On capacity, GPT-5.6 Luna holds more in one request. That matters only if you are near the smaller ceiling — below it, the extra room changes nothing.
- On lifecycle, one of these has a retirement date and the other does not. Requests to a removed model fail outright, so this is the deciding factor regardless of price.
- On long prompts, at least one of these bills the entire request at an elevated rate past a threshold. A comparison made on short prompts can invert once you cross it.
What this page does not tell you. Which model produces better answers. That depends on your task, and this site has no benchmark data — everything above is drawn from published pricing and specifications. Compare quality by running your own prompts through both.
Common questions
- Is GPT-3.5 Turbo or GPT-5.6 Luna cheaper?
- GPT-5.6 Luna is cheaper on input — $0.20 per million against $0.50, a difference of 2.5x. Output rates differ separately, and a workload heavy on generated text can reverse which one costs less overall.
- Which has the larger context window?
- GPT-5.6 Luna, at 1,050,000 tokens against 16,385.
- Can I compare them on my own prompt?
- Yes. The calculator counts any text against every model at once, so you can see both figures for your actual workload rather than for a generic example.
- Do I need to migrate?
- GPT-3.5 Turbo has a retirement date of 2026-10-23. After that the provider removes it and requests fail, so a migration is not optional — only its timing is.