The Short Version
Luna is 5% of Sol's price. On one benchmark it is nearly as good. On the others it is not close. Knowing which of your tasks fall on which side of that line is worth more than any model choice you make this quarter.
Price
| GPT-6 Luna | GPT-6 Sol | Ratio |
|---|
| Input / 1M | $0.10 | $2.00 | 20x |
| Output / 1M | $0.50 | $10.00 | 20x |
| Cached input / 1M | $0.01 | $0.20 | 20x |
Same 1.05M context window and 128K max output on both. The tiers differ on capability and price alone.
Benchmarks
| Benchmark | Luna | Sol | Gap |
|---|
| DeepSWE v1.1 | 66.6% | 68.8% | 2.2pp |
| FrontierCode 1.1 | 42.4% | 49.3% | 6.9pp |
| Agents' Last Exam | 50.9% | 56.4% | 5.5pp |
| OSWorld 2.0 | 52.7% | 64.4% | 11.7pp |
| AutomationBench | 20.7% | 33.2% | 12.5pp |
| Factual error rate | 7.6% | 4.6% | 3.0pp worse |
Reading the Pattern
Sort those rows by gap size and the shape is obvious. The benchmarks where Luna holds up are the ones that hand the model a complete, well-specified problem. The benchmarks where it collapses - AutomationBench at 12.5 points, OSWorld at 11.7 - are the ones that require deciding what to do next, calling a tool, reading the result and adjusting.
Luna executes. It does not plan. That is a clean line and you can route against it.
The Error Rate Matters More Than It Looks
7.6% factual errors against Sol's 4.6% sounds like a three-point difference. In a chain of five dependent steps it is the difference between roughly 68% and 79% of chains completing clean - and every wrong answer in the middle poisons everything downstream.
For anything user-facing without a verification layer, that gap is the reason to pay for Sol.
The Configuration Worth Testing
Luna is cheap enough that you can run it twice on the same input and compare the two outputs, and still spend one tenth of a single Sol call.
Two Luna passes with a disagreement check catches a large share of the errors that make the 7.6% figure alarming, at a price that is still a rounding error next to Sol. Nobody does this yet because nothing has been this cheap before. It is the first thing to try before you conclude you need the bigger model.
Decision Framework
- Classification, extraction, tagging, translation, bulk summarisation - Luna. This is what it is for and the price is not close.
- First-pass code with human or model review after - Luna. 66.6% DeepSWE at 5% of the price is the best value in the current lineup.
- Anything with tool calls in a loop - Sol. The AutomationBench gap is 12.5 points and it widens with every step.
- Browser or computer automation - neither. OSWorld says Astra at 73.5%.
- Customer-facing output with no verification step - Sol. The error rate difference compounds.
- Not sure which bucket a task is in - ask whether the model has to choose its next action. If yes, Sol. If you hand it everything up front, Luna.
Verdict
Luna for volume work where the task arrives fully specified. Sol the moment the model has to make its own decisions about what to do next. The 20x price gap is large enough that the right answer for most teams is both - Luna doing the bulk, Sol handling the judgement - rather than picking one.