GPT-6 Sol and Luna Launch: API Price Cuts, Available in Codex, Not Yet in Chat

OpenAI launched two new models, GPT-6 Sol and GPT-6 Luna, in the early hours of September 23 Taipei time (September 22 in the US). OpenAI says it trained them with similar methods as GPT-6 Astra, bringing some of Astra's advances to faster, more affordable models; GPT-6 Astra remains OpenAI's best model overall, and Sol and Luna extend the lineup downward rather than replacing Astra.

The information here was checked on September 26, 2026; pricing and availability follow what OpenAI's own pages said that day. All benchmark figures were published by OpenAI itself — this site has not tested any of them, and this is not purchasing or investment advice.

Who can use GPT-6 Sol and Luna right now?

Starting the day of the announcement, Plus, Pro, Business, Enterprise, and Edu plans can use GPT-6 Sol and GPT-6 Luna in ChatGPT Work and Codex.

OpenAI's own wording is: "Free and Go users can access GPT-6 Luna in the desktop app." The original text does not specify an operating system, does not say whether it's in Work or Codex, and does not say whether the web or mobile versions can use it; both the announcement and the changelog mention only Luna, not Sol.

OpenAI also writes: "These models are not yet available in Chat." In other words, neither new model is available in Chat yet; OpenAI plans to roll it out in ChatGPT gradually that day, so try again later if you don't see it. Availability also depends on workspace settings, and Enterprise plans need an administrator to enable it first.

On regions and countries, OpenAI hasn't said: none of the four official sources discussing Sol and Luna's availability mention Taiwan, and none list any country or region restrictions either. Whether a Taiwan account can see them can only be confirmed by checking the model menu in the ChatGPT app or Codex directly.

What is the "50% cheaper" compared with: how to read the API pricing

OpenAI's announcement states it is "reducing API prices for Sol and Luna by 50% compared with their GPT-5.6 promotional pricing" — meaning the baseline for this price cut is each of Sol's and Luna's own GPT-5.6 promotional pricing, not the original list price.

The announcement's pricing table marks the price-reduction column "50% cheaper" for both the Sol and Luna rows (see the table below for the exact figures); for Luna's output price, the table prints a drop from $1.20 to $0.50.

The announcement states cached input-token reads get a 90% discount; the pricing page separately lists a cache-write price that is higher than the regular input price — for example, Sol's cache-write price is $2.50 per million tokens, above its regular input price of $2.00, and long context carries higher prices still.

The pricing page states that GPT-5.6 Sol's promotional pricing is available at least through November 21, 2026 — meaning the promotional price Sol is compared against is still in effect today, not an old price; pricing for two generations, GPT-5.6 Sol and GPT-6 Sol, now coexists. The pricing page does not state how long GPT-5.6 Luna's promotional pricing lasts.

OpenAI's own benchmark numbers, compared against Claude

OpenAI says that on its own chosen AutomationBench (a test of business workflows across apps), GPT-6 Sol (xhigh) scores 33.2%, beating Claude Opus 5 (max) at 26.9%, at 11.1 times Sol's cost per task. GPT-6 Luna (high) improves on its own predecessor by 5.4 percentage points, at 58% lower cost per task. The rival models' scores were taken from publicly available reports, and OpenAI itself notes that the table's figure for Claude Fable 5.1 understates its cost, since it omits the cost of Opus 5 fallbacks, which occurred on about 40% of tasks.

OpenAI says that on Agents' Last Exam, Sol (max) scores 56.4%, above Claude Opus 5's highest score on this evaluation, at 60% lower cost; on DeepSWE v1.1, Sol (max) scores 68.8%, within 1.1 percentage points of Fable 5's highest score of 69.9% (xhigh), at roughly 80% lower cost.

OpenAI states these evaluations ran in its own research environment or via its API, and may differ slightly from production ChatGPT; when Claude Fable 5.1 has no score, it reports Fable 5's score instead. For computer use, OpenAI still calls Astra its best model, saying Sol and Luna offer better cost efficiency than their predecessors, not a replacement for Astra.

Caveats on factual accuracy and the safety evaluation

OpenAI's internal factuality evaluation uses real conversations where users had flagged mistakes by its models; on this evaluation, GPT-6 Sol makes about half as many errors as its predecessor, approaching Astra-level reliability at a much lower cost. OpenAI itself states that these error-inducing conversations are not representative of typical usage, where factual errors are rarer.

OpenAI says GPT-6 Luna also improves substantially on factuality: at higher reasoning effort, Luna matches GPT-5.6 Sol at roughly a hundredth of the cost.

Safety information for Sol and Luna is written in an appendix (Section 11) of the GPT-6 Astra System Card, added on September 22, 2026 — not a separate, standalone system card. Under its Preparedness Framework, OpenAI treats both models as High capability in the Cybersecurity and Biological and Chemical domains — its own internal rating, not an outside assessment; neither model reaches the High threshold for AI Self-Improvement, the same as GPT-5.6 Sol/Luna, and OpenAI carries over the same set of safeguards.

The system card appendix also notes that both models hold steady or improve on length-adjusted scores for HealthBench Professional and Consensus, but regress on HealthBench and HealthBench Hard, with noticeably shorter answers: on HealthBench, mean answer length drops by about 45% for Sol and 35% for Luna, with similar reductions on HealthBench Hard.

Want to know if you can use it? Here's how to check yourself

The most direct way to check is to open the ChatGPT app or Codex and see whether GPT-6 Sol and GPT-6 Luna appear in the model menu; OpenAI plans to roll it out on the ChatGPT side gradually that day, so try again later if you don't see it yet.

Developers can call gpt-6-sol and gpt-6-luna directly through the API; in Codex CLI, switch with /model, or start a task with codex --model gpt-6-sol or codex --model gpt-6-luna. For actual billing, use the API pricing page as it reads on the day you make the call.

If your account is on an Enterprise plan, ask your administrator to enable the new models first. None of the four official sources state that GPT-5.6 Sol and Luna are being taken down.

OpenAI's announcement does not mention a rollout timeline for Taiwan or other regions, and does not say whether ChatGPT subscription prices are changing either — API price cuts and subscription fees are two different things; to find out when your own plan gets access, check the actual screen in your own account.

2026 AI News Roundup: Highlights and Daily Life Applications from January to September

GPT-6 Astra Announced: From Answering Questions to Completing Computer Tasks

• AI tools
• Apps & software
• AI news