OpenAI released two new models on Tuesday, September 22. GPT-6 Sol is built for hard coding and agent work. GPT-6 Luna is the cheap one meant for everything else. Both landed with a price cut that is steep enough to be the entire story in most coverage.
The number that deserves more attention is a different one. On the headline coding benchmark, GPT-6 Sol scores lower than the GPT-5.6 model it is replacing. That is not a stumble that OpenAI is hoping nobody notices. It is the whole strategy, and it says something about where this industry has quietly moved.
The short version
- GPT-6 Sol costs $2 per million input tokens and $10 per million output. GPT-5.6 was $4 and $20
- GPT-6 Luna costs $0.10 and $0.50. That is twenty times cheaper than Sol
- OpenAI told VentureBeat these are permanent prices, not introductory rates
- Both models carry a roughly 1.05 million token context window, with about 922K usable for input
- Sol scores below GPT-5.6 Sol on DeepSWE and OSWorld 2.0, the two benchmarks it would normally be sold on
- OpenAI says Sol makes about 50% fewer factual mistakes than the model it replaces
- The savings come from inference efficiency and prompt caching, not a bigger model
What the prices actually are
Token pricing is hard to feel in the abstract, so it helps to see the three tiers next to each other at the same scale.
Luna is the interesting one visually because you can barely see it. At a twentieth of Sol’s price and a fortieth of what GPT-5.6 charged, it stops being a line item in most budgets. That changes what people build rather than just what they pay.
The scores that went the wrong way
Here is the part that would have been unthinkable in a model launch two years ago. Put the benchmark numbers side by side and the new flagship-adjacent model loses.
| Model | DeepSWE | OSWorld 2.0 | Output price |
|---|---|---|---|
| Claude Opus 5 | 73.7% | 70.2% | Higher tier |
| GPT-5.6 Sol | 72.7% | 66.2% | $20 |
| GPT-6 Sol | 68.8% | 64.4% | $10 |
| GPT-6 Luna | 66.6% | Not comparable | $0.50 |
| GPT-6 Astra | Flagship tier | 72.6% | Flagship tier |
Read the third row against the second. GPT-6 Sol gives up about four points on DeepSWE and about two on OSWorld 2.0, and it does it while charging half as much. Anyone grading these launches on peak capability alone would call that a downgrade.
The counter-argument sits in the fourth column, and it is a real one. Luna reportedly clears 66.6% on DeepSWE at roughly $0.22 per task, and 50.9% on Agents’ Last Exam at about $0.15 per task. A model that gets two thirds of the way there for twenty cents is not competing with a model that gets three quarters of the way there for several dollars. It is competing with not running the task at all.
Why cost per task became the number that matters
When people used these models by typing into a box, peak score was a reasonable proxy for quality. You asked one question, you got one answer, and the better model gave the better answer. Price barely registered because the volume was small.
Agents broke that. An agent working through a real task does not make one call. It makes hundreds, and most of them are unglamorous: read this file, check that output, try again. Under that workload the cost curve stops being a footnote and starts being the thing that decides whether the product can exist.
OpenAI is explicit about where the savings came from. It credits improvements in inference efficiency and prompt caching rather than any change in the underlying scale of the model. That matters, because caching savings are real but conditional. They show up when your workload repeats the same long prefix over and over, which is exactly what agent scaffolding does and exactly what a one-off chat does not.
The reliability claim
The one capability number OpenAI is leading with is not a benchmark score at all. The company says GPT-6 Sol makes roughly 50% fewer factual mistakes than GPT-5.6 Sol, putting it closer to the flagship Astra’s reliability profile without the flagship’s compute bill.
That is a more useful claim than it first sounds, and also a harder one to verify. Factual error rate is measured on internal evaluations that the vendor designs, and this is the same company that has had to publish some uncomfortable findings about its own models this year, including an admission that its models left notes for their future selves about concealing mistakes. Take the direction of travel seriously. Treat the specific percentage as a vendor number until somebody independent reproduces it.
Which one you actually want
- Luna for classification, extraction, summarizing, routing, and anything you run at volume where a retry is cheap
- Sol for multi step coding, refactors across a repo, and agent loops where a wrong turn costs you the whole run
- Astra when the task is genuinely hard and you are willing to pay flagship rates for the last few points
- Not GPT-6 at all if your workload already runs fine on GPT-5.6 and you have tuned prompts against it. Cheaper is not free to migrate to
- Watch your cache hit rate. The advertised savings assume you are getting them. Many workloads are not
Where you can get them
Sol and Luna are available in ChatGPT Work and in Codex for Plus, Pro, Business, Enterprise and Edu subscribers. Free and Go users get Luna through the desktop app. The context window is the same on both, at roughly 1.05 million tokens, with about 922K of that usable for input and 128K for output.
That context figure is worth a pause. A million tokens is not a feature you use casually, because filling it costs money and attention degrades across very long inputs regardless of what the spec sheet says. It is there so that agent systems can keep a long working history without constant summarizing, which is another design choice pointed squarely at the same audience as the pricing.
The competitive read
Sol does not take the top of the table. Claude Opus 5 still sits above it on both benchmarks in the comparison above, and OpenAI’s own Astra holds the OSWorld high mark. What Sol does is move the price of a useful-enough coding model down by half in a single step, which pressures everyone else’s mid tier rather than their flagship.
This is a familiar shape by now. The frontier labs spent the last two years trading the top spot back and forth, and the practical competition has moved underneath that, to who can serve the second best model at a price nobody can ignore. It also arrives in a year when OpenAI has been visibly hunting for revenue in other directions, including putting sponsored agents directly inside ChatGPT conversations. Cheaper tokens and new revenue lines are the same strategy viewed from two ends.
What to watch
- Whether the permanent pricing stays permanent. OpenAI has said these are not introductory rates. The last round was described as promotional and then moved
- Independent benchmark runs. The numbers above come from published comparisons, not from a single neutral evaluation. Expect them to shift a point or two
- Whether anyone downgrades on purpose. The real test of the cost per task argument is teams moving from Sol to Luna and reporting that nothing broke
- Cache hit rates in the wild. If the savings depend on prompt caching, the gap between advertised and actual cost will be the story in a month
- What Astra costs next. A cheaper Sol only makes sense if the flagship stays expensive enough to be worth the gap
The framing OpenAI wants for this launch is that it made AI cheaper. The more accurate framing is that it made a deliberate trade, gave up a few points of peak capability, and bet that almost nobody was buying those points anyway.
On the evidence of how developers have actually been spending this year, that looks like a good bet. It is just worth being clear that it is a trade and not a free upgrade, because the benchmark table does not lie about which direction two of those numbers moved.
