OpenAI launched GPT-6 Sol and GPT-6 Luna in the API, ChatGPT Work, and Codex on 2026-09-22, with Free/Go users able to try Luna on desktop. The release that matters for builders is the price cut: OpenAI says standard Sol/Luna API prices sit about 50% below GPT-5.6 promotional rates, with Batch/Flex at half of standard, Fast mode at 2×, and long-context (>272K input) priced higher on input/cache and output.
This post draws on public docs and one independent eval write-up. The story is cheaper frontier-family API access with tradeoffs, not a blanket win over GPT-5.6 or Astra.
What shipped (API)
Model ids are gpt-6-sol and gpt-6-luna. Both advertise a 1,050,000-token context window and 128,000 max output tokens. Published knowledge cutoffs differ: Sol Apr 20, 2026, Luna May 18, 2026.
Standard API prices per 1M tokens (paraphrased from OpenAI model docs):
| Model | Input | Cached input | Cache writes | Output |
|---|---|---|---|---|
| Sol | $2.00 | $0.20 | $2.50 | $10.00 |
| Luna | $0.10 | $0.01 | $0.125 | $0.50 |
Treat the pricing page and model cards as source of truth if numbers move after publish.
Cost efficiency vs quality (Artificial Analysis)
Artificial Analysis (accessed 2026-09-23) argues Sol and Luna push the cost-efficiency frontier more than they rewrite the quality leaderboard.
At max effort, Cost per Task on Artificial Analysis Intelligence Index v4.3 (AA figures):
| Model | GPT-6 Cost / task | GPT-5.6 Cost / task | Approx. drop |
|---|---|---|---|
| Sol | ~$1.06 | ~$1.99 | ~50% |
| Luna | ~$0.07 | ~$0.18 | ~60% |
That tracks OpenAI’s ~50% list-price cut, while both GPT-6 variants used slightly more output tokens per task than their GPT-5.6 counterparts.
Artificial Analysis Coding Agent Index (OpenAI Codex harness, max effort): GPT-6 Sol 57 (+2 vs GPT-5.6 Sol); GPT-6 Luna 41 (−2 vs GPT-5.6 Luna).
Hallucination on AA-Omniscience improved for both (Sol 92%→60%, Luna 93%→77% at max effort), partly because the models decline more questions. Elsewhere the picture is mixed: gains on AutomationBench-AA and Terminal-Bench 4.0 sit beside regressions on GDPval-AA v2.1 (both models) and AA-Briefcase v1.1 (Luna). Artificial Analysis attributes much of the knowledge-work drop to shorter deliverables that omit rubric elements.
Who should care
- If your bottleneck is $/task on agents and coding loops, Sol/Luna are the headline — especially Luna’s price tier.
- If you need a clean quality upgrade over GPT-5.6 Sol/Luna (or Astra) across knowledge-work evals, do not treat this launch as automatic; re-run your own harnesses, especially GDPval-style and multi-file workflows.
- Prefer primary OpenAI model docs and the Artificial Analysis article over secondary roundups (including this one).
The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment