GPT-6 Astra Minor: The Third Name in OpenAI's Azure Leak, and the Only One Nobody Predicted

Three strings turned up in Microsoft's own Azure configuration on September 22, 2026, and only one of them was a surprise. gpt-6-sol and gpt-6-luna had been circulating for two weeks — we covered both, and neither has a model ID, a price or a line in the vendor's documentation today. The third, gpt-6-astra-minor, had never been mentioned by anyone. It is a variant of GPT-6 Astra, the flagship shipped on September 3, 2026, and the name implies something the other two do not: not a new tier beside the flagship, but a smaller version of the flagship itself. GPT-5.6 Luna, the cheap tier already on sale at $0.20 per million input tokens, is the incumbent any cheap GPT-6 name has to justify itself against, and GPT-6 Sol is the sibling whose rumored launch date has already passed once without a product. What follows is what is documented, what is inference, and the one fact about this leak that nobody has reported yet.

What the leak actually is, and what it is not

The claim is narrower than the coverage around it suggests. On the afternoon of September 22, a user on the Chinese developer forum locdd.com posted that Azure's public playground configuration listed three OpenAI registry paths that had not been announced: azureml://registries/azure-openai/models/gpt-6-sol, the same for gpt-6-luna, and azureml://registries/azure-openai/models/gpt-6-astra-minor. The poster gave their source as a live Microsoft URL rather than a screenshot. The Chinese tech press picked it up the same evening and it has been circulating as an "Azure leak" ever since — with a screenshot version of the story now mixed in, which is worth noting because the two versions are not equally checkable.

We fetched the URL the poster named — https://ai.azure.com/modelcache/widgets/PlaygroundConfig.json — on September 23, 2026, and it returns a 200 with a 33 KB JSON document listing registry paths for dozens of models. In the azure-openai registry it names gpt-6-astra and, from the previous generation, gpt-5.6-sol, gpt-5.6-luna and gpt-5.6-terra. It does not contain gpt-6-sol, gpt-6-luna or gpt-6-astra-minor. There are two honest readings of that and we are not going to pick one for you. Either the endpoint is served from a cache that had not picked the new entries up yet, or the three names were seen somewhere the public endpoint does not expose — an internal build, a regional deployment, an authenticated view — and the poster's citation of a public URL is the part of the story that does not hold. Either way, the names themselves remain unattributed to Microsoft or OpenAI, and neither company has said a word about any of them.

It is also worth being precise about what a registry path is. A model registered in Azure's azure-openai registry is a name reservation with a deployment scaffold behind it. It is not a price, not a context window, not a date, not a system card and not a launch. Azure has historically carried model names well before general availability — that is the whole reason this class of leak is worth reading at all, and also the reason it is worth reading sceptically. A name in a registry tells you somebody inside a cloud provider has provisioned for a string. It does not tell you the model exists.

Why astra-minor is the interesting one

The naming grammar is doing real work here, and it is the only real evidence in the story.

In the GPT-5.6 generation, Sol, Terra and Luna are siblings — three distinct names for three tiers of roughly the same family, priced $4/$20, $2/$12 and $0.20/$1.20 per million tokens respectively. A suffix like -minor is a different construction. It attaches to one base model and says this is that model, smaller. It is the grammar of GPT-4o mini, o4-mini, GPT-5-mini, GPT-5.4-mini and gpt-image-2-mini — Microsoft's and OpenAI's own long-standing convention, and the one Microsoft has used across its own small-model line as well.

That distinction matters for what the model would be, not just what it would be called.

• A sibling name like GPT-6 Sol would be a new tier — a different point on the capability-and-price curve, trained and tuned separately.

• A minor name like GPT-6 Astra Minor would be the same generation served cheaper — the same architecture family, a smaller footprint, most likely a distilled or pruned sibling of the flagship rather than a new base.

The second is the more consequential of the two for anyone who writes code against these APIs, because it is the shape OpenAI has used before to make a flagship generation affordable at volume. GPT-4o mini was not a curiosity; it became the default small model across a very large amount of production traffic while the full GPT-4o stayed expensive. If GPT-6 Astra Minor is real and follows that pattern, the interesting number is not its benchmark score — it is whether it lands close enough to Astra's behaviour that you can move a workload down without rewriting the prompts around it.

There is a reason this is speculation and not analysis: nothing about GPT-6 Astra Minor has been published. No price, no context window, no parameter count, no benchmark, no date, no system card. The registry string is the entire evidence base.

What the community thinks it is, and why that is guesswork

The thread where this surfaced is the most useful thing about it, because the guesses are visible and none of them agree. The poster's own summary was blunt: 暂时不清楚是什么定位 — its positioning is unclear. What followed was a set of mutually exclusive theories, and they are worth reading as a sample of how much is actually known, which is nothing.

• A tier between Sol and Astra. The most common guess, and the one the price gap makes most plausible — Astra at $10/$50 is roughly fifty times GPT-5.6 Luna's input rate, and there is a great deal of room in that space.

• A quantised Astra. This one would be genuinely novel: quantisation is normally a community or inference-engine activity on open weights, not something a vendor ships as a first-party registry name. Treat it as the least likely of the four.

• A vision-heavy variant. One reply argued Astra and its minor sibling "read images well but burn tokens and compute" — a claim about a model nobody has run, and therefore not a claim at all.

• A cheap replacement for the previous generation. The argument here is the strongest of the four on its face: Astra is expensive enough that a lighter version "replacing the previous generation is entirely reasonable." That is a statement about OpenAI's incentives, not about the model.

What every one of those theories has in common is that none of them comes with a number. Nobody in the thread claims a spec, because there is no spec to claim. If you see GPT-6 Astra Minor described anywhere with a price, a context window or a benchmark attached, that description was invented.

The gap it would have to fill

It is worth doing the arithmetic, because it explains why a third GPT-6 name is plausible even though nobody has evidence for one.

OpenAI's own published rates put GPT-6 Astra at $10 per million input tokens and $50 per million output, with long-context requests billed at a higher tier above roughly 272,000 tokens. At $50 per million output, a thousand output tokens costs five cents. An agentic loop that emits ten thousand tokens of reasoning, tool calls and diffs costs about fifty cents per run — and that is the cheap part of agentic work, where the loop runs dozens of times per task. Below Astra, the live rates are GPT-5.6 Sol at $4/$20 (on a promotional rate), GPT-5.6 Terra at $2/$12 and GPT-5.6 Luna at $0.20/$1.20. The cliff between the flagship and everything under it is steep, and there is nothing in the GPT-6 generation on the cheap side of it.

That is the commercial case for a minor. It is also the case against the two rumoured siblings: OpenAI already sells GPT-5.6 Luna at $0.20 input, and responding to cost pressure by cutting an existing tier's price 80% — which is what it did in late July 2026 — is cheaper than training and serving a whole new one. A minor is a different proposition, because it is not competing with GPT-5.6 Luna on price so much as competing with it on being GPT-6.

For anyone planning around this, the practical question is not which name wins but how you would try one without betting a production path on it. All four of the OpenAI models that are actually live and routable — GPT-6 Astra, GPT-5.6 Sol, GPT-5.6 Terra and GPT-5.6 Luna — are on OrcaRouter behind a single API key, at the provider's list price with 0% markup passed through rather than marked up, so a vendor rate change is live on our side the same day it lands. That matters more on a tier you call a million times than on one you call a hundred. And if a cheap GPT-6 name does appear, automatic failover is the cheapest way to find out whether it holds up: point a slice of traffic at it, and let the route fall back to the tier you already trust when it does not. To be explicit about what is not true today — nothing called GPT-6 Astra Minor can be routed anywhere, including here, because it does not exist as a callable model. GPT-6 Astra itself is the newest GPT-6 route we carry.

Six days, and a date OpenAI put its own name to

This leak has a natural expiry, and it is short. OpenAI's DevDay 2026 is on September 29 at Fort Mason in San Francisco — announced on April 29, the company's first in-person developer conference since 2024, keynote livestreamed. Sam Altman framed the run-up on September 15 with "a big ship this week, and then for devday ship x 6," walked the first half back the following evening to "next week," and Codex lead Tibo Sottiaux has been teasing the event as the best in company history. The rumoured GPT-6 Sol launch date of Tuesday, September 22 has now passed with no model announced, which is exactly the pattern that pushes attention onto DevDay.

There is one other dated commitment in this story, and unlike everything else in it, OpenAI signed it: the GPT-5.5 series retires on October 14, 2026. Retiring a previous generation on a dated schedule is what a company does when it needs the shelf space, and October 14 is the only date in the entire GPT-6 rumour cycle that the vendor has put its own name to.

So the test is cheap and it is close. If DevDay produces a GPT-6 family, the registry names will look prescient and this piece will be a footnote. If September 29 passes with nothing, that is a real datapoint against all three names at once — and it costs you nothing to wait the six days rather than restructure a roadmap around a string in a config file.

The naming collision that will cost somebody money

Whatever happens on September 29, one trap is already live and it has nothing to do with whether the leak is real.

• GPT-5.6 Sol is a model on sale today. GPT-6 Sol is a rumour. The strings differ by one character and resolve to completely different things.

• GPT-5.6 Luna is a model on sale today at $0.20/$1.20 and is the default for ChatGPT's free and Go tiers. GPT-6 Luna is a rumour with nothing behind it but a naming-symmetry argument from September 6.

• GPT-6 Astra is a model OpenAI shipped on September 3. GPT-6 Astra Minor does not exist, and if it appears it will be a different, smaller model — not a nickname for Astra.

The failure mode is a config file that says gpt-6-luna because a blog post said the tier was coming, and a production job that either errors or silently lands on something else. It is the kind of mistake that is invisible in staging and expensive in production. The discipline that avoids it is boring: wire the exact identifier from the vendor's own model list, and treat any name you have only seen in a leak as a name you cannot call. A 404 from an unannounced model proves nothing about tomorrow — it is the expected answer today.

What would settle it, in order

Four things would turn this from a registry string into a model, and they arrive in a predictable sequence.

• The name appearing in OpenAI's own model documentation or on Azure's models-sold-by-Azure list. That is the first real confirmation and it usually lands before anything else.

• A model ID with a date stamp on it. That is when the model stops being a name and starts being a version.

• The rate card. For a minor, the price is the product. A number meaningfully under Astra's $10/$50 is the whole thesis; a number near it would make the suffix pointless.

• Independent scores. Vendor launch tables for the GPT-6 generation have been generous relative to what independent boards measure, so the first Artificial Analysis placement will be worth more than any launch slide.

Until at least the first two of those exist, the accurate statement is the boring one: three OpenAI model names showed up in an Azure configuration on September 22, 2026, two of them had been rumoured for a fortnight and the third had not been rumoured at all, and neither Microsoft nor OpenAI has confirmed any of them. GPT-6 Astra Minor is the most interesting of the three precisely because it is the one nobody was expecting — a smaller flagship is a different kind of claim than a new tier — and it is also, for now, the one with the least behind it. Six days will tell you more than the last six weeks did.