OpenAI launches GPT-6 Sol and Luna

OpenAI introduced GPT-6 Sol and GPT-6 Luna, expanding the GPT-6 family beyond Astra with models trained using similar methods and API prices about 50% lower than GPT-5.6 Sol and Luna promotional pricing.

SOFTWARE desk — same day as Anthropic’s Opus 5.5 price cut, OpenAI puts cheaper GPT-6 Sol and Luna beside Astra so everyday agent work is not locked to the flagship price.

What the company says the new models are. OpenAI says it trained GPT-6 Sol and GPT-6 Luna with similar methods as GPT-6 Astra. The point of that training, on this page, is to carry the advances behind Astra’s work in professional tasks, factuality, coding, computer use, and alignment into faster, more affordable models. Factuality means getting the facts right. Computer use means the model operates software on a screen, not only replies in text. Sol and Luna are the cheaper way to hand out that GPT-6 generation. Astra, the page says, remains the company’s best model across the board. Choose Astra, it says, when you want the best results and an uncompromising experience. File that ranking as OpenAI’s.

The price, per 1 million tokens. A token is a small chunk of text the model reads or writes. Input is what you send. Output is what the model writes back. The page’s table says GPT-6 Sol is $2 input and $10 output, down from $4 and $20 for GPT-5.6 Sol. GPT-6 Luna is $0.10 input and $0.50 output, down from $0.20 and $1.20 for GPT-5.6 Luna. The table labels both rows 50% cheaper. Sol’s input, Sol’s output, and Luna’s input are exactly half the older promotional rates. Luna’s output is printed as $1.20 down to $0.50, a larger cut than half, and that row is still labeled 50% cheaper. The page says improvements in caching and inference let OpenAI serve the models for less, and that it is passing those savings on by cutting API prices about 50% versus GPT-5.6 promotional pricing. A cache is a saved piece of a prompt you do not pay full price for again. Inference is the work of running the model to produce an answer. The New Stack, same day, adds a line that is not on the launch page: an OpenAI spokesperson told Frederic Lardinois that the GPT-5.6 rates were meant to be promotional, and that these GPT-6 rates are the default price. File the dollar pairs as the launch page’s. File the “default price” sentence as The New Stack’s account of that spokesperson. This desk did not receive an invoice.

Where you can use them. OpenAI says GPT-6 Sol and GPT-6 Luna are available in ChatGPT Work and Codex starting today for Plus, Pro, Business, Enterprise, and Edu users. ChatGPT Work is the work mode inside ChatGPT. Codex is OpenAI’s coding product. Free and Go users can use GPT-6 Luna in the desktop app. Go, on this page, is a plan name, next to Free. The page says the models are not yet available in Chat. In the API, the developer service, the names are gpt-6-sol and gpt-6-luna. To keep service stable, OpenAI says it will roll the models out in ChatGPT gradually through the day. If you do not see them in ChatGPT Work or Codex, the page says to try again later. The New Stack’s availability list names Plus, Pro, Business, and Enterprise, and does not print Edu. The launch page does print Edu. Use the launch page’s list.

Caching, still the company’s, and separate from the sticker cut. OpenAI says it improved prompt caching for GPT-6 so more of a repeated prompt hits the cache by default. Cached input-token reads get a 90% discount. The page also says developers can raise or lower reasoning effort, and turn tools on or off, without throwing away the cached prefix. Reasoning effort is how hard the model works before it answers. A tool is a function the model can call, such as search. Explicit breakpoints let a developer choose where the cached prefix ends. A dashboard and a diagnostics tool show what was cached and what was missed. OpenAI says GitHub reports that, over the past several months, these caching improvements cut the share of prompt tokens that needed fresh processing by more than 50%, across billions of requests to OpenAI models. That GitHub line is a several-month claim OpenAI is citing. It is not a measurement of today’s price cut alone. This desk did not open the dashboard.

The page prints scores. A benchmark is a fixed test used to compare models. OpenAI says its own runs were in a research setup or through the API, which can differ from production ChatGPT, and that competitor scores come from public reports. Where a Claude Fable 5.1 score was not available, the page says it used Claude Fable 5. The cells stay in Sources. This desk did not rerun them. In plain terms, the comparisons OpenAI wants read are about cost per finished task, not only the sticker. On AutomationBench, a Zapier test of business workflows, OpenAI says GPT-6 Sol at its highest effort setting beats Claude Opus 5 at max effort at about 9% of Opus 5’s cost per task, with Sol at 33.2% and a stated cost of $0.27 per task. On DeepSWE, a test of long software-engineering jobs in real codebases, OpenAI says GPT-6 Sol at max effort scores 68.8%, within 1.1 points of Claude Fable 5’s 69.9% at that model’s highest effort, at about 80% lower cost per task. On an offline computer-use test called OSWorld 2.0, OpenAI says GPT-6 Sol at its highest effort scores 60.5%, next to Claude Opus 5 at medium effort at 60.3%, again at about 80% lower cost per task. Treat every one of those figures as OpenAI’s.

What those scores are not. The New Stack says OpenAI did not hand it a full benchmark set ahead of this launch. It also says no one has yet run GPT-6 Sol against Claude Opus 5.5, which Anthropic shipped earlier the same day. Opus 5.5’s sticker, already filed, is $4 per million input tokens and $20 per million output tokens. The New Stack’s arithmetic is that this is still twice GPT-6 Sol’s $2 and $10 sticker, and that Anthropic’s separate “about 40% less than Opus 5 on a typical job” claim is about using fewer tokens, not the sticker. A head-to-head that does not exist is not a ranking this desk made. The New Stack also writes that, as of this story, there is no GPT-6 Terra. That absence is The New Stack’s observation. It is not a sentence on the launch page.

Alignment, in the company’s words, and the limit on the extra percentages. OpenAI says Sol and Luna build on the alignment work introduced with Astra, and that both show improvements over their GPT-5.6 counterparts, including lower rates of misleading claims about their coding work. The page says those evaluations deliberately use hard situations and do not measure how often the models fail in ordinary use. It points to a system card for the full results. This desk did not open that card, and it is not a how-to. The New Stack prints percentages it attributes to OpenAI that are not on the launch page this desk read, including a coding-deception rate for GPT-6 Sol of 1.3%, down from 10.4% for the predecessor. Those figures stay in Sources as The New Stack’s account. A percentage this desk did not read on the primary is not restated here as an OpenAI sentence.

Do not treat a company score as a test this desk ran. Do not read 50% as the Luna output cut: that printed pair is $1.20 to $0.50, while Sol’s rates and Luna’s input rate are exact halves. Do not drop Edu from the paying-plan list. Do not turn “not yet in Chat” into a claim that ChatGPT Work and Codex are off. Do not turn the GitHub caching line into a result of this morning’s price change. Do not invent a GPT-6 Terra denial from OpenAI, a Sol-versus-Opus-5.5 table, or a system-card number this desk did not read.

Plain English for the rest of the card: token = a small chunk of text. input = what you send. output = what the model writes. cache = a saved chunk of a prompt, billed cheaper the next time. inference = running the model to produce an answer. Astra = the GPT-6 model OpenAI still calls its best. Sol and Luna = the cheaper GPT-6 models in this launch. ChatGPT Work = the work mode. Codex = OpenAI’s coding product. API = the service a developer calls. The names there are gpt-6-sol and gpt-6-luna. Go = a plan name on the page, beside Free. benchmark = a fixed comparison test. effort = a setting for how hard the model works. alignment = behavior that stays in line with what people intend. system card = OpenAI’s longer safety write-up. This filing is the 22 Sep launch. It is not an independent score.

PRIMARY here: OpenAI’s 22 Sep 2026 page, “Introducing GPT-6 Sol and Luna,” at openai.com/index/introducing-gpt-6-sol-and-luna/ — Tier A PRIMARY, the company’s own record. The New Stack’s same-day story by Frederic Lardinois is an independent wrap, not a substitute primary. The expansion beyond Astra, the similar-methods line, the professional-work, factuality, coding, computer-use, and alignment carry-over, the $2 / $10 and $0.10 / $0.50 stickers, the $4 / $20 and $0.20 / $1.20 comparison rates, the 50% label, the caching-and-inference savings line, the 90% cached-read discount, the GitHub several-month caching citation, Astra remaining best, the benchmark cells on the page, the ChatGPT Work and Codex availability for Plus, Pro, Business, Enterprise, and Edu, Luna for Free and Go in the desktop app, the not-yet-in-Chat line, the API names gpt-6-sol and gpt-6-luna, and the gradual ChatGPT rollout are OpenAI’s. The spokesperson’s “default price” sentence, the no-full-benchmark-set note, the no Sol-versus-Opus-5.5 head-to-head, the no-GPT-6-Terra observation, the Edu omission, and the alignment percentages that are not on the launch page are The New Stack’s. NOT claimed: an independent rerun, that 50% describes Luna’s output pair, that this desk saw an invoice or the system card, a GPT-6 Terra statement from OpenAI, a stock tip, or investment advice. Distinct from the already-filed claude-opus-5-5, deepseek-v4-1-flash, and openai-frontier-standards.