GPT-6 Sol Leak: Is OpenAI Really Preparing a Sibling to GPT-6 Astra?

There is one problem with GPT-6 Sol as a rumor, and it is not that the model is unannounced — it is that the name is already for sale. OpenAI has been selling a model named Sol since July 9, when GPT-5.6 Sol went live on the API at $4 per million input tokens and $20 per million output, and GPT-5.6 Sol is still very much on the price sheet this week, four days after OpenAI shipped its next-generation flagship GPT-6 Astra at $10 / $50. So when the X account @kimmonismus — the one that surfaced the GPT-5.6 family's Sol, Terra, and Luna names from the Codex app before OpenAI confirmed them — wrote that OpenAI is "already preparing the release of GPT-6 Sol," he was describing something stranger than a simple new launch: a GPT-6 generation that is supposedly a model family, of which the flagship has shipped and the tiered siblings have not.

The claim, posted on September 6, reads: "It appears that OpenAI is already preparing the release of GPT-6 Sol. GPT-6-Astra is their flagship model, and as is known, GPT-6 is a model family; this means Sol, Terra, and Luna are still to come." Then the author trails off into the implications he is curious about. Read the wording twice, because the hedge in the first three words is doing real work: "It appears that" is the language of inference, not of a screenshot or an API listing. No OpenAI announcement, no model identifier, no Codex model-picker capture backs this up yet. It is a thesis from a well-connected observer about where a generation is heading, posted days after GPT-6 Astra's launch and a day after the standard model reached general availability — and it is worth taking seriously exactly because of the pattern it projects.

This is a what-we-know piece, so the labels matter: GPT-6 Sol does not exist yet. Every forward-looking sentence below is a report about what people are saying, not a statement about what OpenAI has shipped. What is measurable sits on the other side of the ledger — what GPT-6 Astra costs, what GPT-5.6 Sol costs, and how OpenAI has historically filled out a generation once its flagship is out.

The leak, and the account behind it

kimmonismus, who runs the AI briefing getsuperintel and posts as Chubby, has a track record that makes this more than a random tweet. His best-known hit is the one this rumor leans on: he published the GPT-5.6 family's retail names — Sol, Terra, Luna — from the Codex application before OpenAI made them public, which is precisely the kind of artifact-based leak that later proved accurate. He has also called Anthropic's release windows closely enough that his September 5 post about a GPT-6 Astra successor, and the same week's report of a Claude Fable 5.2 response, both got relayed as news. None of that makes him right about the future; it makes him worth listening to about what OpenAI is preparing, because his prior leaks had objects in them — strings, screenshots, model pickers. This one, notably, does not.

The other thing to weigh is timing. This GPT-6 Sol post landed about nineteen hours after his successor post, and the two are easy to conflate. Read them as separate claims: the successor post says OpenAI's next flagship, above GPT-6 Astra, is "very soon." This one says the generation below the flagship will eventually fill out with cheaper, tiered models, and that a GPT-6 Sol is being prepared. A lab can do both — ship a more capable flagship and later ship cheaper siblings — but they are different products on different clocks, and treating them as one rumor is how the naming gets tangled.

Where the "family" premise comes from

The load-bearing assumption in the leak is that GPT-6 is a model family. It is worth being precise about where that idea lives, because OpenAI has not actually said it. The company's launch material treated GPT-6 Astra as a flagship — "our most capable model," in OpenAI's phrasing — and did not announce siblings, a roadmap, or a tier structure. The "GPT-6 is a family" framing rests on three things, in increasing order of solidity.

First is the immediate precedent, which is the strongest. The GPT-5.6 generation shipped as exactly this kind of family: GPT-5.6 Sol as the flagship at $4 / $20, with Terra and Luna stepping down in price for the middle and high-volume end of the market. OpenAI did not ship that family all at once either — GPT-5.6 Sol went through a limited, federal-review preview starting June 26 before the full Sol / Terra / Luna set went public on July 9. Astra's rollout has followed the same flagship-first shape: announced September 3, standard model generally available September 5, higher tier still landing on Pro, Business, and Enterprise plans this week. If OpenAI reproduces the 5.6 sequence, siblings follow the flagship. Second is reporting from The Information over the summer that "Astra" was being built as a model family designed for long-running tasks — a class, not a single checkpoint. Third is the leak itself. Two of those three are consistent with kimmonismus's premise; none of them is OpenAI confirming it.

The name already belongs to a model on sale

Here is the wrinkle that makes "GPT-6 Sol" specifically hard to evaluate: OpenAI already sells a GPT-5.6 Sol, and it remains the price-performance pick of the previous generation — live on the API, live on routing platforms at the $4 / $20 list price, with the same ~1.05M-token context window that GPT-6 Astra carries. A reader hearing "GPT-6 Sol is coming" has to decide which of three very different things is meant.

Reading one: a literal sixth-generation tier literally named Sol — OpenAI ships a model called GPT-6 Sol, sold next to a still-available GPT-5.6 Sol with the same celestial name one generation apart. That would be a branding collision OpenAI has never attempted; the 5.5-to-5.6 transition retired its own naming rather than duplicating it. Reading two: "GPT-6 Sol" as shorthand for the sixth generation's workhorse tier, whatever OpenAI ends up calling it — the way the 5.6 family's mid-market model happened to be named Sol, the 6 family's equivalent might carry a different name entirely. Reading three: the rumor is simply leaking upstream, and "Sol" is being used loosely for any cheaper 6-generation model. The article that this rumor points to cannot be written until OpenAI resolves which of these is real, which is itself the most useful thing a reader can take away: the name is doing more work than the model.

What is reasonably clear from kimmonismus's own framing is that he places these tiers below the flagship — "GPT-6-Astra is their flagship model ... Sol, Terra, and Luna are still to come." That inverts the GPT-5.6 hierarchy, where Sol was the top of its generation. Whether a GPT-6 Sol would sit under Astra as a cheaper sibling, or whether the name gets reused at the top of a later generation, is exactly the open question the tweet does not answer.

The price gap a 6-series Sol would close

The reason builders should care about this rumor at all is the price ladder, because that is where OpenAI's family structure has always done its real work. GPT-6 Astra is not expensive by accident: it is $10 in / $50 out per million tokens at the base tier, and requests whose prompts exceed roughly 272K tokens bill at a long-context rate of $20 / $75 on the entire request. GPT-5.6 Sol, by contrast, is $4 / $20, and the tiers below it — Terra, then Luna — step down further toward the high-volume end of the market. A GPT-6 Sol that inherited the 5.6 generation's price slot would be the first way to get sixth-generation intelligence at something like fourth-generation cost, and for teams whose spend is dominated by long, multi-step agent runs, that gap is the whole story.

It is also the part of the leak that has the weakest evidence behind it. No price has been reported. No positioning has been hinted beyond the name. Everything about where a GPT-6 Sol would price is inference from what OpenAI has charged before, which is a reasonable way to bet and a poor way to plan. What the family precedent does establish is direction: OpenAI has repeatedly used the tiered-sibling move to widen the bottom of a generation — the Luna tier of the 5.6 family exists precisely so that high-volume, low-judgment traffic does not have to pay flagship rates.

The strongest case that it is not coming soon

There are two serious reasons to doubt that a GPT-6 Sol arrives quickly, and both are about OpenAI already having built the thing a sibling would be for. The first is reasoning-effort scaling. GPT-6 Astra is not a single price-performance point; it exposes five reasoning levels — low, medium, high, xhigh, and max — and the cheapest settings are where the value math gets interesting. Artificial Analysis measures GPT-6 Astra at medium effort at an Intelligence Index of about 52 for roughly $1.16 per task, a level the tracker places alongside — and slightly above — GPT-5.6 Sol's maximum-reasoning configuration. The cheap setting on the new flagship matching the everything-on setting on the old one, for less money, is the entire reason "set Astra to medium and walk away" has become the working advice of launch week. If the same model already spans the cost range a cheaper sibling would occupy, the business case for shipping a separate GPT-6 Sol is much weaker than the 5.6 precedent suggests — OpenAI may have replaced the tier ladder with a dial.

The second is queue discipline. The same analyst who says GPT-6 Sol is in preparation said a day earlier that a successor above Astra is "very soon," and OpenAI's own executives have spent the week pointing at more capable models on a safety-paced clock. If the next OpenAI release is a new flagship rather than a cheaper sibling, then GPT-6 Sol — if it exists at all — is the release after that, and "already preparing" may describe a model that is months, not weeks, out. None of this disproves the leak. It shifts the prior: the strongest available reading is that OpenAI prepares many things at once, and the first thing to arrive will not necessarily be the one named in a tweet.

What would prove the leak real

Every prior OpenAI leak cycle this blog has tracked left a verifiable object behind before the announcement — the gpt-6-astra identifier surfaced in Codex client code before launch, and the GPT-5.6 family's Sol, Terra, and Luna names leaked from the Codex app's model picker. A GPT-6 Sol that is genuinely in preparation should eventually leave the same kind of trace, and the specific places to watch are few. An identifier in a third-party client or OpenAI's own API changelog — something like a gpt-6-sol or astra-* string — is the strongest signal. A model-picker entry in the Codex or ChatGPT desktop clients is how the 5.6 family leaked and is worth checking whenever a client updates. Cloud catalogs matter too: GPT-6 Astra landed on Azure and AWS Bedrock within days, and a sibling would appear there before any announcement. LMArena codenames, evals catalogues, and Artificial Analysis listings are the lagging indicators — they appear once the model is already being tested or served.

Until one of those objects appears, the honest position is that this is a prediction with a pedigree, not a leak with an artifact. The distinction is not pedantry: it is the difference between the GPT-5.6 naming leak, which had a screenshot, and the countless wrong "OpenAI is about to release X" posts, which had a vibe. kimmonismus's wording — "It appears that OpenAI is already preparing" — sits between the two, and readers should treat it accordingly.

What to do while the family question sorts itself out

None of the decision-relevant facts change on a rumor, because the models that exist are the ones you can call. GPT-6 Astra is live, benchmarked, and available through OpenAI's own API and ChatGPT plans — and on OrcaRouter at OpenAI's list price, $10 / $50, with the long-context tier passed through at $20 / $75, zero markup. GPT-5.6 Sol is live on OrcaRouter as well, at the $4 / $20 price OpenAI set in July, with the same ~1.05M-token context and 128K output ceiling. If you want sixth-generation reasoning today without the flagship's bill, the medium reasoning setting on GPT-6 Astra is the cheapest way to test that claim yourself; if you want the proven workhorse, GPT-5.6 Sol is a call away on the key you already use.

The useful preparation is to make a future GPT-6 Sol cheap to adopt if it arrives, and that is a routing problem rather than a migration problem. When a new tier actually appears in a provider catalogue, a pass-through router makes it callable through the endpoint you already use at the vendor's own list price the same day — no second contract, no rewrite. And because a brand-new sibling is exactly the kind of unproven model you do not want to bet a production path on before it has been load-tested, automatic failover is the honest way to try it: keep the validated model on the primary route, point a slice of real traffic at the newcomer, and let the responses decide. If GPT-6 Sol never ships, nothing about that setup was wasted. If it does, the cost of having believed a tweet is zero.

The fair summary is that a prominent and occasionally accurate OpenAI watcher has pointed at a GPT-6 Sol, and the pattern he is projecting is real even though the model is not yet. OpenAI shipped the GPT-5.6 generation as a Sol / Terra / Luna family after previewing the flagship; GPT-6 Astra has shipped alone; and the price gap between a $10 / $50 flagship and a $4 / $20 previous-generation flagship is exactly the kind of hole OpenAI has historically filled with a cheaper sibling. What is missing is every artifact that would turn the projection into a leak — no identifier, no picker, no cloud listing, no OpenAI word — and what cuts the other way is that a single model with a reasoning dial may already cover the ground a sibling would occupy. The defensible move is to watch the places an identifier would appear, keep calling a model you have validated, and treat the announcement — not the tweet — as the schedule. The name GPT-6 Sol may turn out to be a placeholder, a collision, or the real thing; the family economics behind it are what make the question worth asking at all.