GPT-6 Sol and Luna cut API prices in half

GPT-6 Sol is the model OpenAI is pointing at people who wanted Astra's habits without Astra's bill. The launch post says the pair was trained with methods similar to GPT-6 Astra and is meant to push that generation down the cost curve. TechCrunch's Sept. 22, 2026 story dates the release and adds a competitive detail the post, as opened here, does not: Anthropic had published Opus 5.5 about 90 minutes earlier. This is not investment advice. Prices below are the ones printed on those two pages, not a quote from your dashboard.

TL;DR - OpenAI's table cuts GPT-5.6 Sol promotional pricing from $4 to $2 per million input tokens and from $20 to $10 per million output tokens. Luna goes from $0.20 to $0.10 in, and from $1.20 to $0.50 out. - The post says GPT-6 Sol makes about half as many mistakes as its predecessor on an internal factuality set built from chats where users had flagged errors. - TechCrunch says a gradual rollout to the ChatGPT app and website was expected that day. The launch post says the models are not yet available in Chat.

What actually happened

OpenAI's framing is a family, not a single crown. Astra stays "our best model across the board." Sol is for harder work, including coding. Luna is described, in TechCrunch's paraphrase of the company, as the clerical tier: high-volume tasks with a clear goal, such as summarizing, extracting, or answering quick questions. The post's own line is that work happens at different scales, rhythms, and budgets, and that these two models distribute Astra's gains by being cheaper to serve.

The discount is tied to a specific baseline. The post says API prices for Sol and Luna fall 50% compared with their GPT-5.6 promotional pricing, and it credits caching and inference improvements. TechCrunch says the 6-series models cost half what the 5.6 series cost, and attributes that to the same kind of efficiency work. The word "promotional" is in the OpenAI sentence and not in TechCrunch's summary. If a previous invoice was not on that promotional rate, this article cannot tell you the personal discount. It can tell you the two rows OpenAI printed.

The table is short enough to memorize. Sol: $4 to $2 input, $20 to $10 output. Luna: $0.20 to $0.10 input, $1.20 to $0.50 output. All of those are per million tokens. A fast mode is not given a separate price in the section opened here. Do not invent one.

Factuality is the other headline, and it comes with a caveat in the post itself. On an internal eval built from de-identified real conversations where users flagged a mistake, GPT-6 Sol makes about half as many mistakes as its predecessor and approaches Astra-level reliability at much lower cost. Luna, at higher effort, is said to match GPT-5.6 Sol at about a hundredth of the cost. The post then says those error-inducing chats are not typical usage, where factual errors are rarer, and that scores were not controlled for length, though verbosity sweeps showed almost no dependence on length. TechCrunch quotes the half-as-many-mistakes sentence and the "Astra-level reliability" phrase. It does not dwell on the "not typical usage" caveat. If you repeat the claim, repeat the caveat.

What the benches are actually measuring

A few task prices are specific enough to cite without turning this into a spreadsheet. On AutomationBench, OpenAI says GPT-6 Sol at xhigh effort scores 33.2% at $0.27 per task, above Claude Opus 5 at max effort (26.9%) at 9% of that Opus cost, and above a low-effort Astra score of 30.3% that costs 3.9 times as much per task. A Fable 5.1 line in the same table is 31.4%, with a note that the published cost omits Opus 5 fallbacks on about 40% of tasks, so the true Fable cost is higher than the multiple shown. That footnote is the difference between "cheaper than Fable" as a slogan and "cheaper than a partial bill."

On DeepSWE v1.1, max-effort Sol scores 68.8%, within 1.1 points of Fable 5's 69.9%, at about 80% lower cost per task. Luna at max scores 66.6%. On OSWorld 2.0 offline, Sol at xhigh is 60.5% versus Opus 5 at medium effort at 60.3%, again about 80% cheaper per task. These are OpenAI's comparisons. Production ChatGPT can differ because prompts and tools differ.

Caching is the other discount. The post says GPT-6 prompt caching targets higher default hit rates, with 90% off cached input-token reads, plus a dashboard, diagnostics, and effort or tool changes that keep the cache. GitHub, cited by OpenAI, says these changes cut the share of prompt tokens needing fresh processing by more than 50% across billions of Copilot requests. That 50% is cache hits. The sticker cut is separate.

Where the rollout sentences clash

TechCrunch says the new Sol and Luna models are available in ChatGPT Work and Codex for most paid accounts, and in the API, while Luna is available in the desktop app and for Free and Go users. It then says OpenAI expected to roll the models out to ChatGPT, the app and the website, gradually through the day.

The launch post is stricter. It says Sol and Luna are in ChatGPT Work and Codex starting that day for Plus, Pro, Business, Enterprise, and Edu. Free and Go users can use Luna in the desktop app. It says the models are not yet available in Chat. API IDs are gpt-6-sol and gpt-6-luna. It also says a gradual rollout through the day applies to ChatGPT Work and Codex, so a missing picker means try later. "Not yet in Chat" and "rolling out to the ChatGPT app and website today" cannot both be a full description of the same hour. TechCrunch may be compressing a plan. The post may be describing a gate. Report them as two sentences from two pages, not as one blended status.

TechCrunch says OpenAI claims the pair beats Anthropic's top models, including Fable and Opus. The benches are mixed: higher than Opus 5 at max on AutomationBench, just under Fable 5's best on DeepSWE, roughly tied with medium-effort Opus 5 on OSWorld. "Substantially better" is the posture. The table is the part with numbers. GPT-6 Sol does not win every row it is compared on.

Nearby launches are easy to mash into a single September winner. Gemini 4 Argon is a different company's flagship. Cohere Embed 5 is an embedding model, which none of these coding benches measure. A half-price Sol tier does not tell you who won retrieval, and a retrieval launch does not tell you what gpt-6-sol costs.

What to do as a reader (not a trade)

Copy the price row you will be billed, not the adjective "half." GPT-6 Sol is one row on that menu. GPT-6 Sol at $2 and $10 per million tokens is the promotional-to-new comparison OpenAI printed for the API. Luna's row is $0.10 and $0.50. If someone quotes a different Sol price, ask whether they mean this model, an older promotional rate, or a later revision this article did not open.

If you care about honesty claims, quote the eval the way the post fences it. About half as many flagged mistakes, on a set of chats already known to contain mistakes, approaching Astra, at lower cost. That is not "the model no longer hallucinates." The post says factual errors are rarer in ordinary use than in that set.

If you are wiring an app, use the API IDs and assume Chat, the consumer surface, may lag Work and Codex. Check the picker against the post's "not yet in Chat" line before you write a tutorial that shows a consumer screenshot. And if a benchmark thread drops the fallback footnote on Fable's AutomationBench cost, put the footnote back. A missing 40% of fallback tasks will make a rival look cheaper, or more expensive, than the page you read.

None of this is a reason to trade a stock or a token. The reader version is dull on purpose: match the model, match the surface it is actually on, and match the caveat attached to the brag. The interesting part of GPT-6 Sol is the invoice and the argument about where a mid-tier model is allowed to be wrong. The rest is a leaderboard costume.