GPT-6 Luna Low is the solver default
The solver now uses GPT-6 Luna Low, keeping the Desmos-first strategy library and calculator checks. Explanations and the tutor also use Luna Low by default.
- GPT-5-mini is no longer the active solver or tutor model, and there is no automatic fallback to it.
- Existing bounded correction calls can still use more reasoning within Luna when needed.
- The comparison below helped inform this choice. It is a small diagnostic, not a production speed or accuracy guarantee.
One problem, four configurations
One request per configuration on the same compact, text-only similarity problem. These are API request times, not full SolvePlot solve times.
Scroll the table to compare all columns
| Model | Reasoning | Time | Reasoning tokens | API cost | Answer |
|---|---|---|---|---|---|
| GPT-6.1 Sol | Low | 6.006 s | 168 | $0.009672 | 16Correct |
| GPT-6 LunaCurrent default | Low | 6.819 s | 716 | $0.001072 | 16Correct |
| GPT-6 Luna | Medium | 9.764 s | 1,415 | $0.001735 | 16Correct |
| GPT-5-mini | Low | 17.322 s | 2,304 | $0.011275 | 100Incorrect |
Did the calculator method work?
Sol Low and both Luna configurations produced working proportion regressions for the answer 16. Sol included a readout row. For each one-row Luna fit, an extra parameter readout was added only for the audit to confirm 16; that row was not model-generated.
GPT-5-mini answered 100 and produced an incomplete coordinate construction. Twelve of its fourteen original rows errored in Desmos, so no working method is credited.
Token counts and request settings
All four requests used 463 input tokens, 0 cached tokens, a 4,096-token output cap, Low verbosity, JSON mode, and requested Priority processing. There were no retries or separate explanation calls.
| Configuration | Visible output | Total output | Original rows |
|---|---|---|---|
| GPT-6.1 Sol Low | 223 | 391 | 2 · no errors |
| GPT-6 Luna Low | 263 | 979 | 1 · no errors |
| GPT-6 Luna Medium | 227 | 1,642 | 1 · no errors |
| GPT-5-mini Low | 770 | 3,074 | 14 · 12 errors |
Total output includes reasoning tokens. Sol and Luna returned the service-tier label Fast; Mini returned priority. Costs use the returned usage and applied tier.
What this comparison can tell us
Luna Low found the same useful proportion at a lower request cost in this test. A single attempt does not establish typical speed, accuracy, or reliability. The compact test excluded the full strategy library, production candidate selection, image input, and separate explanation generation.
Mini ran after the other three requests. Timing was not controlled for provider load, and no p50 or p95 is inferred. The listed costs are usage-derived API request costs, not customer prices or invoice totals.