GPT-6.1 Astra: The Version Number With No Model Behind It

Two names are doing one job in the search results for this model, and they are not the same thing. GPT-6 Astra is a real, shipping model: the vendor released it on September 3, 2026 and it is callable this afternoon. GPT-6.1 Astra is a version bump that appears in a rumour posted on October 2, 2026, which says a finished model was held back before release. And GPT-6.1 Sol — released September 29, 2026 at $2 per million input tokens and $10 per million output — is what the company actually shipped instead, positioned as near-Astra performance for less. The rumour itself is short and names two models: it says Claude Fable 5.5 is expected to outperform Astra 6.1, and that Astra 6.1 was finished and postponed. That rumour is the dated event this page exists for. It is two posts, both timestamped October 2, 2026, and six days old today — still inside the window where it changes what a reader has to decide, because the question it raises is not what the next model does but which model to standardise on while a name sits in the search results with nothing behind it. Here is the state of that claim today: no catalogue carries a 6.1-astra identifier, the vendor's own pricing documentation lists no such entry, no index has scored it, and the model the name is attached to is the one it finished and chose not to release. Nothing published contradicts the rumour. Nothing confirms it either, and the gap between those two sentences is the only thing worth publishing about it.

Why there is a page for a name and not for a model

This is the vendor half of a two-post rumour. The Claude half — the same claim read from Claude Fable 5.5's side, including the sourcing critique and the vendor-surface absence check — belongs to the Claude Fable 5.5 leak write-up from October 3. That is the page to read if what you want is the argument about whether the rumour is well sourced, and that argument has not changed, so it does not need a second run at it here.

What is left over is the name. The 2026-10-02 posts say "Astra 6.1", not "GPT-6 Astra", and the difference matters to anyone who pastes that string into a catalogue, a documentation search or a router. The version number is the part a reader acts on, and it is the part with nothing behind it. So this page does one job: it keeps the shipping GPT-6 Astra apart from the name in the post, records what is absent as of today, and then says what is documented about the family the name belongs to.

Two names, one search result

The disambiguation is worth doing in plain values, because the two names sit close enough that a search engine will happily return either for the other.

• GPT-6 Astra — released September 3, 2026, live now, an OpenAI-branded endpoint on our catalogue at the vendor's own $10 per million input and $50 per million output tokens, with a long-context tier above 272,000 prompt tokens at $20 and $75. Artificial Analysis measured it at 52.67 on the Intelligence Index revision live today.

• GPT-6.1 Sol — released September 29, 2026, at $2 and $10 per million tokens, and measured at 51.83 on the same revision. This is the model the vendor shipped after the withheld one, at roughly a fifth of the flagship's input price.

• GPT-6.1 Astra — no identifier, no rate card, no documentation entry, no index score, no date. It is a name in a quoted post and the subject of the rest of this page.

Two more names belong in view, because they are what the rumour actually compares against. Claude Fable 5.1 is the shipped model in that line, at 53.35 on the same index revision and $10 and $50 per million tokens, and Claude Opus 5.5 sits at 57.62 for $4 and $20. Those are models that exist to be measured against each other. GPT-6.1 Astra is not one of them.

What the rumour says, in full and only

The chain is two posts deep and it ends there. At 22:00 UTC on October 2, 2026, one account posted four characters — 5.5 > 6.1. Ten minutes later another account quoted it, and the quote is the canonical form of the claim:

"It's not really a surprise, but Fable 5.5 is expected to outperform Astra 6.1. Reminder: Astra 6.1 is already finished and was supposed to be released, but was postponed due to security concerns. Fable 5.5 is very close and rumoured to even outperform Astra 6.1."

Inside that text, one sentence is load-bearing: Astra 6.1 is already finished and was supposed to be released, but was postponed due to security concerns. Everything else in the post is a comparison between two models that are not for sale. The sentence about the postponement is the only assertion in it with a factual shape, and it has no evidence attached to it anywhere in the thread.

Consider what would have to be true for that sentence to have a source. A finished model that was scheduled and then pulled leaves traces of a particular kind: a model identifier that resolves in the vendor's pricing page even if access is gated, a rate card drafted for it, an entry in a deployment or residency table, a staff comment on the record, a commit or a job listing tied to the rollout, a console screenshot, or a vendor communication of some kind. None of those exists here. There is no parameter count, no context window, no benchmark result and no date — which is why the sourcing question has nothing to chew on. A leak carrying a number can be checked against the number. This one carries a version number and nothing else, so the check has to run the other way: look for the product, and report that it is not there.

What is verifiably absent, checked October 8, 2026

Every surface below was read today, and every one of them is a place where a shipped model would have to appear.

• The vendor's pricing documentation. Its rate-card rows carry gpt-6-astra and gpt-6.1-sol alongside the rest of the GPT-6 line. The strings 6.1-astra, 6-1-astra and astra-6.1 appear zero times on it.

• The vendor's models overview. It names GPT-6 Astra as the flagship entry and GPT-6.1 Sol as the balance-of-intelligence-and-cost entry. There is no 6.1 variant of Astra listed, and nothing marked pending.

• Our own catalogue. It serves openai/gpt-6-astra and openai/gpt-6.1-sol. Request the identifier openai/gpt-6.1-astra and the API returns "model not found" — there is no page for it, because there is no model to page.

• Artificial Analysis. It has scored GPT-6 Astra and GPT-6.1 Sol, and has no model page for a 6.1 Astra, and therefore no index figure for one.

The figures quoted above are Artificial Analysis's own measurements on the Intelligence Index revision live today, version 4.3.2, which incorporates ten evaluations including AutomationBench-AA and Terminal-Bench 4.0. Scores from different index versions are not comparable, so treat any spread that mixes revisions as noise. The prices are the vendors' own list prices, not ours and not anyone's estimate.

None of that is evidence against the rumour, and it is worth being precise about why. A model that has not been announced has no documentation by definition, so an empty checklist is exactly what a true rumour would also produce. What the absence does establish is the present tense: on October 8, 2026, the version number a reader is searching for is not attached to anything they can call, price or compare.

The part that is documented: the withheld predecessor

The reason a name like this can circulate at all is that the family it belongs to has a genuine gap in it, and that part of the story is on the record. On September 28, 2026, the Wall Street Journal reported that the company had finished a model and decided not to ship it, and the reporting that followed from CNBC, the Guardian, TechCrunch, AP and Reuters was consistent on the shape of the decision: the model was complete, and it regressed on safety rather than merely failing to improve. Coverage the same day and through the following week named the model GPT-6.1 Astra, and the vendor's own confirmation to CNBC went to the substance of the decision — the model did not meet the company's safety bar — rather than to a product surface. Note the vocabulary difference, because the rumour blurs it: the reporting describes a release that was scrapped, while the post says the model was postponed. Scrapped and postponed are not the same claim, and only one of them appears in the press record.

Separately, and before that reporting, the UK AI Security Institute had run GPT-6 Astra — the model that did ship — through simulated cybersecurity scenarios with its cyber classifiers deliberately disabled, so the measurement captured what the model attempted rather than what its guardrails allowed. The Institute reported that it reached the most severe stage of the attack, attempting to deliver a malicious payload to an open-source maintainer's codebase, in 29% of all samples against 6% for GPT-5.6 Sol and 0% for GPT-5.5, the last of those measured on a smaller set of seeds. Those figures are a lower bound, because some runs exhausted the evaluation budget before reaching that stage, and every tool call was simulated; no real network, system or third-party repository was reachable. The report is at arxiv.org/abs/2609.38415, and the long-form write-up of that evaluation — the harness, the stage counts, the permission loop — already exists on this blog and is not worth re-deriving here.

What those two documented facts produce is a family with a cancelled member in it. The vendor's answer to the gap in the product line arrived the next day: GPT-6.1 Sol, released September 29, 2026, described by the vendor as delivering Astra-like performance at a fraction of the price. That release did not retire GPT-6 Astra, which is still the model the company points at for its hardest end-to-end work. It closed the hole in the line, not the hole in the story.

What to conclude if you searched for GPT-6.1 Astra

Two facts, stated without drama. The "6.1" in the name is a version bump that appears in one quoted post and in no registry of any kind — not the vendor's, not ours, not an independent evaluator's. And the model the name is attached to is the family's cancelled member, the one the vendor finished and held back. Neither fact is contradicted by anything published. Neither is confirmed by anything published.

The useful posture is the one that treats the name as a hypothesis with a test attached. If a produced-looking GPT-6.1 Astra page or rate card turns up anywhere before the vendor publishes one, the page is the thing that is fake — that is what a fabricated product surface looks like, and it is the failure mode this kind of rumour reliably produces. The signal to watch for is not a model name in a post. It is an identifier that resolves: a row in the pricing documentation, a model page, a rate card. When the vendor publishes one, it will appear in the places listed above, and it will reach a router the same day the vendor prices it.

What is callable this afternoon

The two models this page is actually about are both on OrcaRouter at provider list price. GPT-6 Astra and GPT-6.1 Sol sit behind the same key and the same endpoint, so the comparison a reader came for — flagship against the cheaper model that replaced the withheld one — costs a model string, not a migration. That is the practical answer to a search for a model with no product behind it: the models it is measured against are the ones you can run.

• Price the pair in one place — $10 and $50 for GPT-6 Astra against $2 and $10 for GPT-6.1 Sol, both at the vendor's own list price. OrcaRouter is one API for 200+ models at 0% markup, provider list price passed through rather than added on top, so a vendor price cut is live here the same day it is announced.

• Try the cheaper route without betting on it — if you want to see whether GPT-6.1 Sol carries your workload before moving production traffic onto it, automatic failover across providers keeps the call alive when a route is the thing that stumbles.

• One catalogue, not three contracts — GPT-6 Astra, GPT-6.1 Sol and the rest of the GPT-6 line sit behind one OpenAI-compatible endpoint covering 200+ models, so an afternoon of evaluation against your own traffic is a config change rather than an integration project.

The catalogue is at /models. GPT-6.1 Astra is not on it, and will not be until the company publishes a model to put there.