GPT-6 Luna and GPT-6 Sol: The Cheap Tier OpenAI Hasn't Shipped

There is one hard, checkable artefact in this story, and it is not either of the names in the headline. On September 21, developers reported seeing the string GPT-6 Sol medium inside NVIDIA code merge records — a third party's repository, not the vendor's. Around that single artefact a much larger rumour has assembled: that OpenAI is preparing to launch GPT-6 Sol as a mid-tier workhorse and GPT-6 Luna as its cheap tier, below the flagship GPT-6 Astra that OpenAI actually shipped on September 3, 2026. Neither GPT-6 Sol nor GPT-6 Luna has a model identifier, a price, a context window, a system card or a line in OpenAI's own documentation. The tier they would occupy, though, already has a live occupant in GPT-5.6 Luna — the budget member of the GPT-5.6 family, on sale since July 9, 2026 at $0.20 per million input tokens and $1.20 per million output. That asymmetry is the whole story, and it is the reason this piece is worth reading rather than merely worth hedging.

What is actually documented, and what is inference

Start with the artefact, because it is the only thing here that a reader can go and look at. The NVIDIA merge record is real in the sense that the people reporting it were reading a public repository, and the code change described is mundane: a commit written with the assistance of GPT-6 Astra and then reviewed by something labelled GPT-6 Sol. The 新智元 write-up that carried the finding to a wider audience framed it as a working example of an "AI 打工矩阵" — a chain where one model writes and another reviews. Treat that framing as commentary. The merge record itself tells you somebody inside a third party's infrastructure is preparing to serve a model under that name. It does not tell you OpenAI has shipped one, and it says nothing at all about Luna.

The rest of the case for a GPT-6 Sol launch is thinner than the coverage around it suggests. Chinese tech press has reported a leaked price of $2.50 per million input tokens and $15 per million output — roughly half of what GPT-5.6 Sol costs — alongside claims that support was being wired up across the OpenAI API, Codex OAuth, image input and long-context handling, and a claimed launch on Tuesday, September 22. Some of that reporting pointed at 3 a.m. Tuesday. At the time of writing there is no OpenAI announcement, no identifier string, no pricing-page entry and no model-picker sighting for GPT-6 Sol or GPT-6 Luna. One outlet covering the same rumour chain noted that its own price figures did not agree with each other, citing $2.50/$15, $4/$30 and other numbers in circulation — which is what a rumour looks like after it has been through four translations, not what a price list looks like.

GPT-6 Luna is weaker still, and it is worth being blunt about that. Luna's presence in the story traces to a single post from September 6, in which the X account @kimmonismus wrote that OpenAI was "already preparing the release of GPT-6 Sol" and then reasoned from the GPT-5.6 naming pattern that "Sol, Terra, and Luna are still to come." That is an argument from naming symmetry. Luna has never been the subject of a claim in its own right — no leak, no screenshot, no model string, no price. It appears in later roundups only because it appeared in the first post.

Three dates, and what each one proved

Leak timelines are worth reading as a scorecard, because this one already has one.

• September 6 — the family claim. Community inference from the GPT-5.6 precedent, no artefact attached. Unverified then, unverified now.

• September 15–16 — Sam Altman posted that OpenAI had a "big ship this week, and then for devday ship x 6," then walked the first half back the following evening, saying the thing he was most excited about launching would arrive the next week instead and would be "worth the wait." He named no product on either post. This is the only executive-stated element in the entire story, and it speaks to timing and nothing else.

• September 21 — the NVIDIA merge record, and the first artefact in the chain. A merge record in someone else's repository is a real signal that a model is being prepared for serving. It is not a release.

• Tuesday, September 22 — today, and the date the rumour itself named. As of this writing there is no OpenAI announcement, no model identifier, no price-list entry and no model-picker sighting for GPT-6 Sol, and none at all for GPT-6 Luna.

The next dated checkpoint is DevDay on Tuesday, September 29 at Fort Mason Center — the "ship x 6" Altman referred to, and the fallback date the Chinese reporting pointed at if the Tuesday launch did not happen. If a sixth-generation mid-tier or budget model is real, that is where it most plausibly surfaces, and a single line in OpenAI's own model list would settle everything this article has to hedge.

The tier already exists, and it is cheap

Here is the part of the story that is entirely checkable, and it is the reason a "GPT-6 Luna" is a strange thing to wait for. GPT-5.6 Luna is live, and it is cheap in a way frontier models are not.

• Price — $0.20 per million input tokens and $1.20 per million output, per OpenAI's own API documentation. It launched on July 9, 2026 at $1.00 / $6.00 and was cut by roughly 80% in late July.

• Cached and long-context rates — cached input reads run at roughly a 90% discount; requests past the long-context threshold are billed at about $0.40 / $1.80 for the whole request.

• Context and output — a 1M-token context window with a 128K maximum output. Median time to first token is listed at 1.83 seconds.

• Role — OpenAI replaced GPT-5.5 Instant with GPT-5.6 Luna as the default model for ChatGPT Free and Go on August 6–7, 2026, adding a "Think" button for deeper reasoning the following week.

• Effect of the cut — on September 9, OpenAI's chief financial officer Sarah Friar told the Goldman Sachs Communacopia conference that the price reduction produced roughly a tenfold increase in usage, and argued that deploying Luna is cheaper than running Z.ai's GLM 5.3 on a cloud layer. Both are vendor figures delivered by an executive at an investor event: worth reporting, not worth treating as independently verified.

Put the two price points side by side and the tiering question answers itself. The reported GPT-6 Sol price of $2.50 / $15 is more than twelve times GPT-5.6 Luna's input rate. A sixth-generation Luna would not be competing with Sol — it would be competing with the model OpenAI is already giving away as its free-tier default.

Why a cheap GPT-6 tier is plausible anyway

There is a real commercial argument underneath the rumour, and it is about where the tokens are rather than where the leaderboard is. GPT-6 Astra costs $10 per million input tokens and $50 per million output, with long-context requests billed at a higher tier. It lists a 1M-token context window — the same class as the 5.6 generation, at a much higher rate. On Artificial Analysis, the max-reasoning configuration of Astra scores 53 on the tracker's Intelligence Index at about $3.26 per task, which is a defensible price for frontier work and an indefensible one for classification, extraction, routing or high-volume agent loops.

That gap is exactly the shape that produced Sol, Terra and Luna in the 5.6 generation, and it is why the reporting keeps pointing at a mid-tier Sol rather than a second flagship. OpenAI's own pricing behaviour supports the direction of travel: its response to cost pressure in July was to cut the cheap tier by 80% rather than to replace it, and its CFO went on stage in September to argue about price against Chinese open-weight models. A company defending the low end of the market is a company that intends to keep shipping at the low end.

The counter-argument is equally real, and it is why the leak's own logic is weaker than it looks. Shipping a 6-series budget model means asking every team that has already tuned prompts, caching and evaluation harnesses around GPT-5.6 Luna to re-qualify all of it for a marginal gain. That is a genuine reason a GPT-6 Luna might never be announced — not because the rumour is false, but because the product decision it assumes may not have been made.

How to prepare without betting on a model that does not exist

You cannot call GPT-6 Sol or GPT-6 Luna. Nobody can, and no platform hosts them — the most that has been reported is that one developer saw a GPT-6 Sol model string returned where a GPT-5.6 Sol request was expected, which is a routing curiosity rather than availability. What you can do is make the eventual arrival cheap to adopt and expensive to be surprised by.

The practical version is to keep the model name out of your application code. If your stack calls a model by a hard-coded string, adopting a sixth-generation mid-tier means a deploy, a prompt re-qualification pass and a new evaluation run. If it calls a route, it means changing one line. OrcaRouter exposes 200+ models behind a single API with 0% markup — provider list price passed straight through, so when a vendor cuts a price the number on our side moves the same day, which is how GPT-5.6 Luna's 80% cut showed up here without anyone renegotiating anything. Automatic failover does the rest: point a route at the cheapest tier that satisfies your eval threshold, and an unshipped model stays out of the path until it exists and passes.

That is the honest use of a leak like this one: not to plan a migration to a model that has not been announced, but to notice which slot in your stack a new mid-tier would fill, and to make sure filling it later is a config change rather than a project.

What would count as evidence

The line between a leak worth following and a leak worth ignoring is whether the artefact is something the vendor or a neutral party produced. Applied here:

• Counts — a model identifier in OpenAI's API documentation, a system card, an entry on OpenAI's pricing page, an appearance in the ChatGPT model picker, a first independent benchmark run on a tracker like Artificial Analysis, or a listing in a routing catalogue with a real endpoint behind it.

• Does not count — a merge record in someone else's repository on its own, an unattributed rumour list in a roundup, a leaked screenshot of a price that three outlets report three different ways, an unnamed developer's observation, or a prediction-market probability. Prediction markets were pricing roughly 77–78% for a GPT-6 release by September 30 back in late July. That is a market's opinion about a date, not evidence about a model, and it was priced before Astra shipped.

Until one of the first list lands, the accurate position on both names is narrow and unglamorous. GPT-6 Sol is a string in an NVIDIA merge record and a price nobody can agree on. GPT-6 Luna is a name that appears in a family rumour, in a list nobody has claimed, occupying a tier that GPT-5.6 Luna already fills at $0.20 in and $1.20 out. There is nothing to migrate to, nothing to benchmark, and nothing to price. The one thing worth doing today is making sure that when there is, your routing does not need rewriting to reach it.