GPT-6 Sol
- Provider
- OpenAI
- Status
- Current
- Context
- 872,000 tok
- Price
- $2 / $10 /MTok
GPT-6 Sol and GPT-6 Luna are OpenAI’s mid-weight and small models, released on 22 September 2026 below GPT-6 Astra (TechCrunch). They succeed GPT-5.6 Sol and Luna at half the price: Sol at $2/$10 per million tokens against $4/$20, and Luna at $0.10/$0.50 against $0.20/$1.20. OpenAI credits improved caching and inference and bills both as standard rates, not introductory ones.
Artificial Analysis measured the generation as a cost move rather than a capability one: Sol scores 48 on the Intelligence Index v4.3 at max effort against GPT-5.6 Sol’s 47, and Luna 37, level with its predecessor, while cost per completed task halves — $1.06 for Sol and $0.07 for Luna. Neither model has a published SWE-bench figure.
Quick specs
| GPT-6 Sol | GPT-6 Luna | |
|---|---|---|
| Released | 22 September 2026 | 22 September 2026 |
| Input price | $2.00 / MTok | $0.10 / MTok |
| Output price | $10.00 / MTok | $0.50 / MTok |
| Previous generation | $4 / $20 | $0.20 / $1.20 |
| Context window | 872,000 tokens | 1,000,000 tokens |
| AA Intelligence Index | 48 (max) | 37 (max) |
| AA Coding Agent Index | 57 (up 2) | 41 (down 2) |
| Cost per AA index task | $1.06 | $0.07 |
| Output speed | ~131 tokens/s at max | 140–163 tokens/s |
| Hallucination rate (max) | 60%, from 92% | 77%, from 93% |
What changed
The generation is a price move. Artificial Analysis’s summary is that index and Coding Agent Index scores stay level with GPT-5.6, with progress on some evaluations and regressions on others, while pricing falls by half or more.
Prices. Sol $4/$20 to $2/$10; Luna $0.20/$1.20 to $0.10/$0.50. OpenAI attributes the cut to better caching and inference rather than a smaller model.
Cost per completed task. $1.99 to $1.06 for Sol, and $0.18 to $0.07 for Luna, on Artificial Analysis’s measurements. That is the lowest cost per completed index task among full-size frontier-adjacent models on our board.
Factuality. OpenAI says Sol makes about half as many mistakes as its predecessor on factuality evaluations. Artificial Analysis measured hallucination rates at max effort falling from 92% to 60% for Sol and from 93% to 77% for Luna — an improvement from a high base.
Regressions. Both models score lower than their predecessors on GDPval-AA v2.1, and Luna drops about 45 Elo on AA-Briefcase v1.1 and two points on the Coding Agent Index.
Benchmark performance
| Measure | GPT-6 Sol | GPT-5.6 Sol | GPT-6 Luna | GPT-5.6 Luna |
|---|---|---|---|---|
| Intelligence Index v4.3 (max) | 48 | 47 | 37 | 37 |
| Coding Agent Index (max) | 57 | 55 | 41 | 43 |
| Cost per index task | $1.06 | $1.99 | $0.07 | $0.18 |
| Output tokens per task | ~31,000 | ~29,000 | ~51,000 | ~41,000 |
| Hallucination rate (max) | 60% | 92% | 77% | 93% |
Source: Artificial Analysis. By effort, Sol reads 48 at max, 43 at high, 40 at medium and 28 without reasoning; Luna reads 37 at max down to 18 without reasoning.
How they compare
vs GPT-6 Astra
GPT-6 Astra scores 53 on the index at $10/$50 and $3.26 per task. Sol is five points behind at a fifth of the token price and a third of the cost per task. OpenAI describes Sol as reaching Astra-level reliability at much lower cost; on the index it does not reach Astra’s score.
vs Claude Opus 5.5
Claude Opus 5.5 leads the index at 58 and costs $4/$20, with $5.98 per completed task. Sol is ten index points behind at roughly a sixth of the cost per task.
vs Grok 4.7
Grok 4.7 scores 46 at $2/$6 per million tokens. Sol is two points ahead on the index and, despite a higher output price, costs less per completed task — $1.06 against $2.73 at high effort — because Grok 4.7 spends far more output tokens.
Known limitations
No SWE-bench figure. Neither model publishes SWE-bench Pro or Verified, and no standardised harness has produced one.
Capability is flat. The index moves one point for Sol and none for Luna; the generation’s gains are in price.
Regressions on knowledge work. Both drop on GDPval-AA v2.1, and Luna also falls on AA-Briefcase and the Coding Agent Index.
Hallucination rates remain high. Artificial Analysis measured 60% for Sol and 77% for Luna at max effort on AA-Omniscience, improved but still high in absolute terms.
Frequently asked questions
When were GPT-6 Sol and Luna released?
22 September 2026, on the OpenAI API and in ChatGPT Work and Codex for most paid accounts, with Luna also rolling out to the desktop app and to Free and Go users.
How much do GPT-6 Sol and Luna cost?
GPT-6 Sol is $2 per million input tokens and $10 per million output tokens, half GPT-5.6 Sol’s $4/$20. GPT-6 Luna is $0.10/$0.50, against $0.20/$1.20 for GPT-5.6 Luna. OpenAI bills both as standard prices rather than introductory rates.
Is GPT-6 Sol better than GPT-5.6 Sol?
Marginally on capability and substantially on price. Artificial Analysis scores it 48 on the Intelligence Index against 47, and 57 on the Coding Agent Index against 55, while cost per completed task falls from $1.99 to $1.06. It regresses on GDPval-AA v2.1 knowledge work, and neither model publishes a SWE-bench figure.
Where does GPT-6 Sol rank?
It enters our best AI models ranking at #8, above GPT-5.6 Sol, and takes the board’s best-value slot on cost per completed task.
Written at launch, 23 September 2026. Artificial Analysis figures are independent. OpenAI’s own launch page was not reachable at the time of writing, so prices and availability are taken from Artificial Analysis and TechCrunch’s launch coverage.