OpenAI releases GPT-6 Sol and Luna at half the API price of GPT-5.6
Sol costs $2/$10 and Luna $0.10/$0.50 per million tokens, half the price of the GPT-5.6 versions. Artificial Analysis scores them 48 and 38, one point up on their predecessors, with a regression in office-style work; Sol was replaced a week later.
What happened
On 22 September OpenAI released GPT-6 Sol and GPT-6 Luna. According to TechCrunch, the announcement came about 90 minutes after Anthropic released Claude Opus 5.5.
OpenAI’s pitch: GPT-6 Astra, released three weeks earlier, started a new generation, and these two models make that intelligence “more efficient and accessible”. Sol is the workhorse for coding and multi-step work. Luna is for “high-volume tasks with a clear goal”, such as summarising documents, extracting information or answering quick questions.
Where you could use them on launch day:
- ChatGPT Work and Codex (ChatGPT’s agent workspace and coding tool) for Plus, Pro, Business, Enterprise and Edu users;
- Free and Go users could try Luna in the desktop app only;
- The API, as
gpt-6-solandgpt-6-luna. GitHub Copilot added both models the same day.
The ordinary chat window was not included. Only on 7 October did ChatGPT switch its default chat model to GPT-6, with Sol for paying users and Luna for Free and Go users; see ChatGPT’s interactive answers.
| Tier | Model | API price (per million tokens, input / output) | Best for (OpenAI’s description) |
|---|---|---|---|
| Top | GPT-6 Astra | $10 / $50 | The hardest end-to-end work across steps and tools |
| Workhorse | GPT-6.1 Sol (from 29 Sept) | $2 / $10 | Repeated, long-running complex work |
| Workhorse | GPT-6 Sol | $2 / $10 | The same (now superseded by 6.1 Sol) |
| Middle | GPT-5.6 Terra (previous generation) | $2 / $12 | No GPT-6 version |
| Economy | GPT-6 Luna | $0.10 / $0.50 | Extraction, classification, rewriting, structured summaries |
Sources: OpenAI API pricing and ChatGPT model documentation, checked 9 October 2026. Prompts longer than 272,000 tokens are billed at higher rates.
How big is it? The same ability for half the money
The headline is price. OpenAI says it made caching and inference more efficient and is passing the savings on:
- Sol API price (GPT-5.6 Sol: $4 / $20)
- $2 / $10
- Luna API price (GPT-5.6 Luna: $0.20 / $1.20)
- $0.10 / $0.50
- Discount on cached input
- 90%
- Context window (tokens)
- 1.05M
Sources: OpenAI API docs; Artificial Analysis, 22 September 2026. Prices per million tokens.
Did capability get discounted too? Artificial Analysis, an independent benchmarking firm, runs every model through the same ten tests (coding, knowledge, science, agent tasks) and combines them into one score. Its finding: GPT-6 Sol and Luna score about the same as their predecessors, one point higher each.
Claude models tested with Anthropic's default fallback enabled; Gemini 4 Argon at its highest (high) setting. Source: Artificial Analysis model pages, checked 9 October 2026
The real difference is what each answer costs. By Artificial Analysis’s measurements, one task from its index costs:
Source: Artificial Analysis, article of 22 September 2026 and model pages checked 9 October 2026 (GPT-6 Sol was $1.06 at launch)
Put another way: $6 buys one task from Opus 5.5, five or six from GPT-6 Sol and more than 80 from Luna. Opus 5.5 is 10 points ahead of Sol, and Sol is 10 points ahead of Luna. That is what the extra money buys.
Looking at individual tests, Artificial Analysis also found:
- Less making things up. When Sol didn’t know an answer, it invented one 60% of the time, down from 92%; Luna fell from 93% to 77%. Sol gets there by declining more often: it attempts 83% of questions, against 99% for GPT-5.6 Sol, so its share of correct answers also fell, from 59% to 54%.
- Better at command-line work, but far from the top. On Terminal-Bench 4.0, Sol rose from 40% to 44%; Luna scored 13%. Opus 5.5 and GPT-6 Astra score about 60% on the same test.
- Worse at office deliverables. On GDPval-AA, which simulates real work across 44 occupations, Sol fell by about 100 Elo points and Luna by about 75. The deliverables were shorter and more often left out required elements.
GPT-6 Sol was tested before OpenAI fixed a bug affecting image understanding; Artificial Analysis expects little impact. Source: Artificial Analysis scores as quoted in Anthropic's Claude Opus 5.5 and Claude Sonnet 5.5 announcements
What OpenAI says
The main claims in OpenAI’s announcement, all from its own testing:
- Half the mistakes. On an internal test built from real conversations where users flagged errors, “GPT-6 Sol makes about half as many mistakes as its predecessor, reaching Astra-level reliability at much lower cost.”
- Less misleading talk about its own code. OpenAI measures how often a model makes deceptive claims about coding work it has done:
Source: OpenAI launch chart, as reposted in the OpenAI Developer Community announcement, 22 September 2026
- Business workflows for less. On Zapier’s AutomationBench, which asks an AI to carry out real business processes across connected apps, OpenAI’s chart shows Sol peaking at about 33%, against about 29% for GPT-5.6 Sol, at less than half the cost per task. Astra scores 41.4%.
- Plainer answers. Both models adopt Astra’s communication style: less jargon, fewer low-value details and slightly shorter answers.
According to MacRumors, OpenAI compared Sol with Anthropic’s Opus 5 and Fable 5.1 and said it matched or beat Fable 5.1 on some tests. It did not compare against Opus 5.5, released the same day.
How it compares
On price, the four big labs’ mid-range models now sit almost on top of each other:
| Model | Released | API price (per million tokens, input / output) | Intelligence Index |
|---|---|---|---|
| GPT-6 Sol | 22 Sept | $2 / $10 | 48 |
| Claude Sonnet 5.5 | 28 Sept | $2 / $10 | 56 |
| Gemini 4 Argon | 30 Sept | $2 / $10 (launch discount; list $4 / $20) | 53 |
| Claude Opus 5.5 | 22 Sept | $4 / $20 | 58 |
| GPT-6 Luna | 22 Sept | $0.10 / $0.50 | 38 |
| Claude Haiku 5.5 | 7 Oct | $0.10 / $0.50 (prompts up to 100,000 tokens) | 43 |
Sources: official vendor pricing; Intelligence Index from Artificial Analysis, 9 October 2026.
The same per-token price does not mean the same bill. Artificial Analysis measured Sonnet 5.5 at its maximum setting using about 193,000 output tokens per task, roughly seven times GPT-6 Astra, so it costs far more per task than Sol. At its “high” setting, though, Sonnet 5.5 lands almost exactly where GPT-6 Sol does on both score and cost. Luna’s rival, Haiku 5.5, scores five points higher but uses about three times as many tokens per task at maximum settings.
Things to keep in mind
- “Half the mistakes” and the deception rates are OpenAI’s own numbers. Independent tests confirm the price cut and the drop in made-up answers, but also found a regression in office-style work that OpenAI did not mention.
- It was the main model for one week. On 29 September OpenAI released GPT-6.1 Sol at the same price with a score four points higher, and its documentation now recommends 6.1 Sol. There are few reasons to choose GPT-6 Sol directly today.
- The savings come from the price cut, not from using fewer tokens. Both new models actually use slightly more output tokens per task (Sol 31,000 vs 29,000; Luna 51,000 vs 41,000).
- An index is an average. On individual tests the rankings shift, and different benchmarkers use different settings.
What it means for you
- ChatGPT Free and Go ($8 a month) users: since 8 October your chats use GPT-6 Luna by default. There is nothing to choose.
- Plus ($20 a month) and above: chat defaults to GPT-6 Sol; in Work and Codex, OpenAI recommends GPT-6.1 Sol. By OpenAI’s estimate, a Plus user can send roughly 15 to 150 Sol messages in Codex every five hours, or 350 to 3,000 with Luna.
- Old models are going: GPT-5.5 leaves ChatGPT, Work and Codex on 14 October (the API is unaffected). Any saved setting that uses it should move to Sol or Luna.
- Developers: as a rough YLEM estimate, condensing a 20-page report (about 15,000 tokens) into a 1,000-token summary costs about 0.2 cents with Luna, 4 cents with Sol and 20 cents with Astra, not counting reasoning tokens. Batch processing halves that. High-volume jobs with clear rules are where Luna pays off.
For live rankings see our models page; for ChatGPT’s plans see the ChatGPT product page.