FULL STORY

GPT-6 Sol and Luna: From Leak to Launch in Three Days

An Azure configuration leak on Sept 21 revealed GPT-6's dual models Sol and Luna, fueling speculation before Luna surfaced early in Codex. On Sept 23, OpenAI officially launched both models at half the price.

2026-09-21 ~ 2026-09-23 · 4 episodes · 93 posts

Episode 1 · Azure configs leak GPT-6 Sol, Luna and Astra Minor, fueling OpenAI release rumors (2026-09-21, 33 posts)

On September 22, anonymous leaker scaling01 posted that OpenAI's next-generation models GPT-6-Luna and GPT-6-Sol are "coming soon," hinting OpenAI may adopt a dual-version naming scheme; the account later added "Azure had a little mishap," leading observers to speculate the model names and release plans leaked early via the Azure platform. The account also casually noted Qwen4 is quite close.

Confirmed

  • All of the above comes from the leaker account and media retellings; there is no officially confirmed information from OpenAI so far.

Unconfirmed

  • A leak relayed by danielmac8 claims OpenAI is in the final preparation stage for releasing GPT-6 Sol and Luna, with the Terra tier to be dropped; a roundup by 新智元, however, says GPT-6 Sol may come alongside Terra and Luna as a trio — the two accounts differ on whether Terra stays.
  • Per the 新智元 roundup, GPT-6 Sol is the mid-tier model, faster with lower token consumption; leaked pricing is $2.50/$15 (input/output), half of GPT-5.6 Sol, with a rumored Tuesday release; NVIDIA foundry details were also mentioned in the leak (not elaborated in the source material).
  • The leak relayed by danielmac8 says Anthropic is also working on version upgrades involving Fable, Opus, and Sonnet 5.5, with Opus 5.5 launching at a discount — read as a move to undercut OpenAI.

Why it matters

  • If the price-cut leak holds, GPT-6 Sol would significantly lower the unit cost of high-end models and could trigger a new price war; simultaneous upgrade rumors around Anthropic and OpenAI suggest frontier-model competition is entering a dense iteration cycle. Note that all information comes from anonymous accounts with questionable reliability, so readers should treat it with caution.

13 more related posts →

Episode 2 · Codex Leak Hints at GPT-6 Luna, Sol Still Missing (2026-09-22, 2 posts)

Users spotted a "GPT-6 Luna" model selectable in OpenAI Codex while "Sol" is absent; earlier code leaks showed gpt-6-luna answering requests with pricing built in, suggesting a staged rollout.

Episode 3 · OpenAI Rolls Out GPT-6 Sol and GPT-6 Luna (2026-09-23, 10 posts)

On September 23, multiple leakers and users reported that OpenAI is gradually rolling out two new GPT-6 series models, GPT-6 Sol and GPT-6 Luna. An official announcement is believed to be imminent, but as of publication OpenAI has not formally confirmed the news, and specific specs and benchmark data remain unavailable.

Confirmed

  • The tipster account TestingCatalog said OpenAI is pushing GPT-6 Sol and GPT-6 Luna to ChatGPT, Codex and the API, with screenshots attached.
  • Blogger markk and Reddit user ees-h tested and found both models already usable in OpenAI Codex and Work, with corresponding options appearing in the model selection interface.
  • Developer danielmac8 said both models appeared in his Codex model selector and predicted an official announcement was coming soon.
  • User kimmonismus (relayed by airesearch12) said they could already use both models and promised to publish benchmark data once obtained.
  • AI community insider Andrew Curran (relayed by rickasaurus) said OpenAI was launching both models that day, with Sol and Luna rolling out and already available to some users.

Not Yet Confirmed

  • There is no official information on the specific specs, capability differences, pricing or benchmark results of GPT-6 Sol and Luna, and no test data has leaked from the posts (kimmonismus said they will publish once they get it).
  • The dual codename "Sol (sun) and Luna (moon)" appears only in m5's account; whether OpenAI will adopt it as the official naming remains to be seen.

Why It Matters

  • This would be a rare "double launch day" for OpenAI—releasing two flagship models on the same day, which several leakers see as a pattern similar to its past release style. If true, it would mark the GPT-6 series debuting in a dual-model form, directly affecting the ChatGPT, Codex and API ecosystems, making official announcements and first benchmarks worth watching closely.

Episode 4 · OpenAI Launches GPT-6 Sol and Luna at Half the Price (2026-09-23, 48 posts)

On September 23, OpenAI officially released two new models in the GPT-6 family, GPT-6 Sol and GPT-6 Luna. They went live the same day in ChatGPT Work and Codex (covering Plus/Pro/Business/Enterprise/Edu users) and opened up via API (gpt-6-sol and more). Built on the technical foundation of GPT-6 Astra, the two models target faster, cheaper performance for large-scale workloads. Sam Altman made the announcement personally, calling them the next generation of the GPT series and a powerful shake-up of the AI model pricing landscape.

Confirmed

  • Sam Altman's announcement: Sol and Luna deliver significant gains over the 5.6 series in intelligence, alignment, work output, coding, and computer operation, at half the per-token price and lower per-task compute cost; the OpenAI developer account confirmed API pricing is 50% below GPT-5.6, with the positioning "build with Sol, scale with Luna."
  • Specific pricing: per kimmonismus's roundup, Sol costs $2 input / $10 output, Luna $0.10 input / $0.50 output per million tokens; polynoamial noted this is 50% below even GPT-5.6's promotional pricing.
  • Availability: OpenAI announced Free and Go users can use Luna in the desktop app.
  • Performance data: per reachvb and official figures, Sol (xhigh) beats competitors on AutomationBench at 9% of the cost; Sol scores 48.4% on FrontierCode (xhigh); GPT-6 Luna Max matches GPT-6 Sol High and Astra Low on the DeepSWE benchmark at roughly 1/3 of Sol's price; officially, task costs on some coding evaluations are up to 93% cheaper than Claude Opus 5.

Unconfirmed

  • Polymarket cautions that comparisons like "93% cheaper than Claude Opus 5" come from official claims and still await third-party benchmark verification.
  • Pricing screenshots posted by some users (e.g., yeshvvanth) should be checked against OpenAI's original announcements.

Why it matters

  • Sam Altman added that measured by per-task pricing, GPT-6 Sol and Luna "have no competition in the market," and he wants users to consume heavily; the community sees the aggressive pricing as pressure on the entire competitive landscape.
  • Users like indiantruely believe the pricing "could bury plenty of open and closed models"; ns123abc mocked things up with a "Sir" meme noting Sol's pricing is half of Claude Opus 5.5's — flagship models going head-to-head the same day, with a price war now the talk of AI circles.

28 more related posts →