GPT-6 Sol and Luna: what launched and what it costs
OpenAI shipped GPT-6 Sol and Luna on 22 September 2026: Sol for coding and agent work at $2/$10 per million tokens, Luna for bulk clerical jobs at $0.10/$0.50. What changed, and what it means for a content team.
OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September 2026, three weeks after GPT-6 Astra. Sol is the mid-tier model aimed at coding and agent work; Luna is the cheap one for high-volume clerical jobs. Both cost about half what the GPT-5.6 versions did, and neither generates images, video or audio. TechCrunch, 22 September 2026 has the launch; the prices below are from OpenAI's API pricing page.
What launched
Three GPT-6 models now sit side by side:
- GPT-6 Astra (3 September 2026): the flagship. Strongest, slowest to reach every plan, and the most expensive.
- GPT-6 Sol (22 September 2026): OpenAI describes it as the model for "complex tasks like coding". TechCrunch reports that Sol "makes about half as many mistakes as its predecessor, reaching Astra-level reliability" on factual accuracy.
- GPT-6 Luna (22 September 2026): for "clerical work — high-volume tasks with a clear goal, like summarizing documents, extracting information, or answering quick questions".
The API model IDs are gpt-6-astra, gpt-6-sol and gpt-6-luna.
What it costs
List prices per million tokens, from OpenAI's pricing page at launch. OpenAI prices a short and a long context tier; it does not state where the boundary sits.
| Model | Input (short) | Cached input | Output (short) | Input (long) | Output (long) |
|---|---|---|---|---|---|
| GPT-6 Astra | $10.00 | $1.00 | $50.00 | $20.00 | $75.00 |
| GPT-6 Sol | $2.00 | $0.20 | $10.00 | $4.00 | $15.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 | $0.20 | $0.75 |
Two ratios are worth holding on to. Sol is a fifth of Astra's price on both input and output. Luna is a twentieth of Sol's. TechCrunch puts the pair at "half the cost of the 5.6 series".
For comparison, Anthropic's Claude Opus 5.5, released the same day, lists at $4 input and $20 output per million tokens, so Sol is half its price at list. List price is not cost per task, though: a model that needs fewer attempts or fewer tokens to finish can be cheaper at a higher rate. See our Opus 5.5 write-up for its numbers.
The benchmarks OpenAI and the press reported
VentureBeat's launch coverage lists these figures:
- AutomationBench 1.0.6 (Zapier's workflow benchmark): Sol 33.2%, at $0.27 per task.
- DeepSWE 1.1: Sol 68.8%, Luna 66.6%.
- OSWorld 2.0 (operating a desktop): Sol 60.5%.
- Agents' Last Exam: Sol 56.4%.
- Internal deception test: Sol 1.3%, down from 10.4% for GPT-5.6 Sol.
Read these the way you read any launch-day table. They are vendor-selected, run by or for the vendor, and the versions matter: an AutomationBench score on 1.0.6 is not directly comparable with a score on a different release of the same suite. The deception figure is OpenAI's own internal test, not an external audit.
The Luna number is the surprising one. On DeepSWE it sits 2.2 points behind Sol at a twentieth of the price. For the jobs Luna is aimed at, that gap may not matter at all.
Where you can use them
According to VentureBeat, Sol and Luna are rolling out in ChatGPT Work and Codex for Plus, Pro, Business and Enterprise accounts, and in the API. Free and Go users get Luna in the desktop app. At launch neither was in the regular ChatGPT chat window yet; OpenAI said the ChatGPT rollout would be gradual.
Early write-ups, such as Jake Handy's model notes, put the context window at about 1.05 million tokens with up to 128,000 tokens of output. OpenAI's pricing page does not state it, so treat that as reported rather than confirmed.
What it means for a content team
Neither model makes a frame of video, an image or a voice track. Nothing in the launch mentions media output. If your week is short-form video, Sol and Luna change the text and agent half of the work, not the rendering half.
That half is still most of the hours. Here is where each one fits.
Luna: the bulk jobs you used to skip. Luna is priced for work you run hundreds of times: summarising call transcripts into post ideas, pulling product claims out of a catalogue, tagging a back library of clips by topic, first-pass subtitle clean-up. At $0.10 per million input tokens, reading a million tokens of transcripts costs ten cents at list. The jobs that were not worth a model call at GPT-5.6 prices are worth one now.
Sol: the agent that drives the pipeline. Sol is the reasonable default for an agent that calls tools, reads errors and fixes its own mistakes: the loop that picks a model, submits a generation, polls it and writes the result back. Its claimed factual-accuracy gain matters most here, because an agent that invents a model name or a duration wastes a paid render.
Astra: keep it for the hard, rare problems. At five times Sol's price, Astra is for the tasks Sol fails, not the daily loop. We covered where it helps in GPT-6 Astra will not make your videos.
How this fits with Versely
Versely's agent writes the brief, picks the video, image or voice model, prices the job in credits and makes it. The chat model behind the agent is a choice in the model picker, and that list now includes Claude Opus 5.5. GPT-6 Sol and Luna are not in it today.
The split still holds whichever language model you prefer: the language model plans and calls tools, and a media model such as Veo 3.1, Kling or Seedance makes the frames. Changing the planner does not change what a render costs. To try the agent side, open the Versely agent or create a free account.
FAQ
When did GPT-6 Sol and Luna launch?
22 September 2026, three weeks after GPT-6 Astra (3 September 2026).
How much does GPT-6 Sol cost?
$2.00 per million input tokens and $10.00 per million output tokens in the short-context tier, with cached input at $0.20. Long context is $4.00 in and $15.00 out.
How much does GPT-6 Luna cost?
$0.10 per million input tokens and $0.50 per million output tokens in the short-context tier; $0.20 and $0.75 for long context.
Is GPT-6 Sol better than GPT-6 Astra?
No. OpenAI positions Astra as the flagship and Sol below it; VentureBeat reports Sol is less capable than Astra. Sol's case is price: it is a fifth of Astra's cost, and OpenAI says it reaches Astra-level factual reliability.
Can GPT-6 Sol make videos or images?
Nothing in the launch mentions image, video or audio output. For video you still need a video model; a language model can plan the job and call one.
Is GPT-6 Sol available for free?
Not at launch. Free and Go users can use Luna in the ChatGPT desktop app; Sol is for Plus, Pro, Business and Enterprise accounts and the paid API.