DataLLM LabGPT Astra: The Model Is Confirmed, the Name Is Not
OpenAI has confirmed Astra. It calls the system its next major model, and announced it by publishing ten previously unsolved results in mathematics and theoretical computer science produced by an internal version. What OpenAI has not confirmed is the thing most articles put in their headline: whether it ships as GPT-6. Reporting says the name is undecided between GPT-6, something like GPT-5.7, or a fourth model class beside Sol, Terra and Luna, and that even “Astra” is tentative. There is no release date. This page sorts what OpenAI said from what is circulating, then does the one concrete thing available: we have measured Sol, Terra and Luna, so here is what the classes Astra would join actually cost and score.
Astra is unusual among unreleased models: OpenAI announced it itself, with a concrete artefact attached. That makes the confirmed layer larger than usual and the rumour layer easier to separate.
The ledger: confirmed, attributed, unsourced
| Claim | Tier | Where it comes from |
|---|---|---|
| Astra exists and is OpenAI's next major model | Confirmed | OpenAI's own description |
| An internal Astra produced ten previously unsolved results in mathematics and theoretical computer science | Confirmed | OpenAI published them as the announcement |
| Designed for long-running work — plan, revise, delegate and keep working in the background for hours or days | Confirmed | OpenAI's description of the system |
| Multiple agents collaborating over long periods on hard problems | Confirmed | Same |
| Sam Altman demonstrated it to policymakers and regulators in Washington | Attributed | Reported by The Information |
| Altman described watching it use a computer in a super-human, very fast way, and framed the result as persistent agents | Attributed | Same reporting |
| The name is undecided — GPT-6, GPT-5.7, or a separate class beside Sol, Terra and Luna. “Astra” itself is tentative | Attributed | Same reporting |
| The model family is already in testing | Attributed | Same reporting |
An ultima-alpha partner early-access programme | Unsourced | No official confirmation we could find |
An internal checkpoint named mozaik-alpha-fdm | Unsourced | No named source |
| A 3–10 September rollout window | Unsourced | Forecast, not announcement |
| A linked GPT-Image 2 update | Unsourced | No named source |
| Running continuously for weeks | Unsourced | OpenAI's own framing is hours or days |
Notice which tier the specifics fall into. The codenames, the dates and the week-long autonomy — the details that make coverage feel like inside knowledge — are the ones with nobody's name attached. The capability claims, unusually, are OpenAI's own.
Why GPT-6 is the wrong headline
If you searched for GPT-6 and landed here, the useful answer is: the model is real, the name is not.
Reporting is explicit that OpenAI has not decided whether Astra ships as GPT-6, as a point release in the GPT-5 line, or as a new class alongside its existing Sol, Terra and Luna families. All three of those already exist and ship today, which is what makes the third option credible rather than a hedge — OpenAI has been naming model classes rather than incrementing a version number for a while now.
So an article confidently titled “GPT-6 release date” is guessing at two things at once: a date nobody has given, and a name OpenAI says it has not chosen. We made the same point about a different vendor in the Grok 5 page, where the widely repeated parameter count turned out to have no named source at all.
Sol, Terra and Luna, measured
Here is the one concrete thing available today. Astra is reported as joining the class system that Sol, Terra and Luna already occupy. We have run all three on our executed Python benchmark, so the shape of that lineup is measurable even though Astra is not.
| Model | Score | Measured cost / 1k tasks | Latency | Reasoning tokens |
|---|---|---|---|---|
| GPT-5.6 Luna Pro | 8/9 | $3.79 | 7.2s | 548 |
| GPT-5.6 Sol | 9/9 | $4.98 | 6.6s | 58 |
| GPT-5.6 Terra Pro | 9/9 | $6.99 | 5s | 246 |
| and for scale, elsewhere in the same lineup | ||||
| GPT-5.4 mini | 9/9 | $0.53 | 2.3s | 0 |
| GPT-5 | 9/9 | $11.00 | 22.9s | 924 |
Two things fall out of that table. The classes are not a quality ladder — Luna Pro is the cheapest of the three and the only one that dropped a task, while Terra Pro is the most expensive and the fastest. And the whole named-class tier sits between $3.79 and $6.99 per 1,000 tasks, against $0.53 for GPT-5.4 mini on the same nine tasks with a clean 9 out of 9. Whatever Astra is priced at, that is the neighbourhood it would be born into.
What a persistent agent would cost
The confirmed capability claim is the economically interesting one: a model that plans, delegates and keeps working in the background for hours or days.
Every hour of that is tokens. On our benchmark, per-call reasoning spend across models that all scored 9 out of 9 ranges from 0 tokens to 1,327, and reasoning bills at the output rate. A single call's difference is pennies. Multiplied across a session that runs for a day and coordinates several agents, it is the entire invoice — the arithmetic we worked through in our agent model comparison and again in what an AI agent actually costs.
So the number to watch when Astra ships is not its benchmark score. It is tokens per hour of autonomous work, which no vendor publishes and which decides whether a persistent agent is a tool or a subscription you cannot afford to leave running. We will measure what we can when it is callable.
What nobody can tell you
- The release date. There is none. Any date you see is a forecast.
- The name. GPT-6, GPT-5.7 and a separate class are all live options per the reporting, and “Astra” is itself tentative.
- The price or the benchmarks. Nothing outside OpenAI has been run. Any comparison table showing Astra against other models is projection.
- What we did not do: we have not run Astra, seen it, or been briefed on it. The confirmed tier above is OpenAI's public description; the attributed tier is us reading named outlets.
When Astra is callable it goes through the same nine tasks as every other model here, and we publish whatever comes back.
How our numbers were produced
Nine Python tasks, each a function signature plus a spec and no example tests. Generated code executes against hidden asserts in an isolated python3 -I subprocess with a 12-second timeout. Temperature 0, max_tokens 4000, one scored attempt per task. Cost is derived from measured token counts at the list price on each model's measurement date, not a billing statement, and prices move. Runs go through OpenRouter. Full method on the methodology page.
FAQ
Is Astra the same as GPT-6? Undecided. Reporting says OpenAI has not chosen between GPT-6, a GPT-5 point release, and a separate class alongside Sol, Terra and Luna.
What has OpenAI actually confirmed about Astra? That it is their next major model, that an internal version produced ten previously unsolved results in mathematics and theoretical computer science, and that it is built for long-running multi-agent work over hours or days.
When is Astra coming out? No announced date. A 3–10 September window has circulated with no official source.
What is ultima-alpha? A codename reported for a partner early-access programme. We found no official confirmation that the programme exists.
How much will Astra cost? Unknown. For context, the existing named classes it would join measured between $3.79 and $6.99 per 1,000 tasks on our benchmark.