AIREITEROpenAI says Astra will be available “soon,” but it has not announced a ChatGPT or API release date. This status was checked against OpenAI’s public updates on September 3, 2026.
The verified Astra timeline
OpenAI’s official record shows research disclosure, safety review, and controlled-release preparation—not a public launch.
| Date | What OpenAI’s public record says | What it does not establish |
|---|---|---|
| August 1, 2026 | OpenAI published its mathematics and theoretical-computer-science announcement describing results from an internal Astra version. | It was not a ChatGPT or API launch. |
| August 7, 2026 | OpenAI said preliminary evaluations could not rule out the Critical cybersecurity capability threshold for Astra. | It did not publish a consumer release date or price. |
| August 18, 2026 | OpenAI said it had paused two weeks of deployment-focused reinforcement learning and kept its largest planned frontier run on hold while safeguards were strengthened. | It did not say Astra was canceled. |
| August 28, 2026 | OpenAI later reported restarting a previously paused large frontier reinforcement-learning run. | Restarting training did not equal public availability. |
| September 1, 2026 | OpenAI said Astra meets its Critical cybersecurity threshold, that safeguards were sufficient to proceed under its framework, and that it plans to make Astra available soon. | “Soon” is not a launch date, ChatGPT plan, API promise, or price. |
The primary sources are OpenAI’s August 7 cyber-capability update, August 18 development update, and September 1 Path to Astra announcement.
What “release” means for ChatGPT Astra
OpenAI has announced a planned release, but no public ChatGPT model-picker entry or developer endpoint is documented in the reviewed OpenAI material as of September 3.
| Access question | Verified status as of September 3, 2026 |
|---|---|
| Can ordinary users select Astra in ChatGPT? | No public confirmation in the reviewed OpenAI updates. |
| Is there a public Astra API? | No official endpoint, model ID, or API documentation in the reviewed OpenAI updates. |
| Is Astra available through ChatGPT Plus, Pro, Team, or Enterprise? | No plan-by-plan access terms in the reviewed OpenAI updates. |
| Is there a public price? | No Astra subscription or token price in the reviewed OpenAI updates. |
| Are any testers expected to receive access? | OpenAI says the most advanced cybersecurity capabilities will initially be limited to a small group of testers. |
| Will defensive users get a broader route? | OpenAI says access for defensive cybersecurity work is planned through Daybreak Blue; no public application route is described in the reviewed announcement. |
| Is a system card available? | OpenAI says more safety and security details are expected in a system card at launch. |
Treat “available soon” as a status update, not a live product, until OpenAI publishes access documentation.
What OpenAI has actually confirmed about Astra
OpenAI’s public safety material publishes no conventional spec sheet for Astra: no confirmed context window, parameter count, latency target, license, or general-purpose benchmark table. The company describes Astra as an upcoming model with advances in coding and cybersecurity capability.
Under OpenAI’s Preparedness Framework, the Critical cybersecurity threshold concerns a model that can identify and develop functional zero-day exploits across many hardened, real-world critical systems without human intervention, or devise and execute novel end-to-end cyberattack strategies from a high-level objective. OpenAI says Astra meets that threshold after internal evaluations and expert-led assessments.
The most concrete figures in the September 1 announcement are evaluation figures, not consumer performance promises:
| Reported measure | Astra result or status | Important qualification |
|---|---|---|
| ExploitBench | 100% | OpenAI’s benchmark for developing exploits from known vulnerabilities. |
| Internal evaluation set | 20 high-severity vulnerabilities | Created to reduce contamination concerns, according to OpenAI. |
| Zero-day discoveries | 2 | Reported during internal evaluation and used in an exploit chain. |
| Cyber-jailbreak refusal rate | 91.5% | OpenAI’s evaluation of disallowed cyber-assistance requests. |
| Advanced cyber access | Restricted initially | Results were obtained with Daybreak Blue access, not described as default public behavior. |
OpenAI says Astra is more token-efficient than GPT-5.6 Sol on its internal cybersecurity evaluation. The announcement does not provide complete arbitrary-code-execution percentages, token counts, or a full reproducible benchmark table, so this is not an independent leaderboard comparison.
The mathematics announcement is evidence, not an access announcement
OpenAI’s mathematics announcement says an internal Astra version produced ten advances across mathematics and theoretical computer science. The openai/ten-proofs repository contains Lean 4 formalizations intended to make the arguments machine-checkable.
Lean verifies the formalized steps, not the significance, originality, human contribution, or general usefulness of the research. The announcement is evidence of a research workflow, not a public Astra trial.
The widely repeated $2,000 number is not Astra pricing. It refers to the estimated token cost of successful research runs at GPT-5.6 Sol API rates, according to the mathematics announcement and reporting around it; it excludes product pricing, failed attempts, training costs, and a full accounting of human research labor.
Why the launch has been slower than rumor dates
OpenAI says Astra’s cybersecurity capability required stronger controls before development and deployment could proceed at full speed.
The August 7 update described isolated testing, restricted network and tool access, model-weight protection, encryption, sandboxed execution, and broader monitoring. The August 18 update added that OpenAI had temporarily slowed reinforcement-learning work while it hardened research environments and expanded alignment testing.
The September 1 announcement changed the status from “Critical cannot be ruled out” to “meets the Critical threshold,” but it did not remove the controls. OpenAI said safeguards were sufficient to proceed while advanced cyber capabilities remained restricted at first.
OpenAI’s September 1 update warns that additional monitoring can slow, pause, or stop legitimate work when an agent has tools or handles cybersecurity-related tasks. The system card at launch should explain those boundaries in more detail.
OpenAI Astra is unrelated to Google’s Project Astra, the camera-and-voice assistant research project that Google announced in 2024 and later integrated into Gemini features. The shared name does not indicate a shared model or product.
How to evaluate the next Astra announcement
If a new Astra date appears on social media, check it against these four items before treating it as real:
- An OpenAI source with a publication date. A post from a leaker, prediction market, or third-party AI directory is not a launch notice.
- A usable product surface. Look for an Astra entry in ChatGPT, an official API model ID, or documented access for a named tester program.
- Safety and capability documentation. The promised system card should clarify which capabilities are available, restricted, or monitored.
- Commercial terms. A real public release should specify eligible plans, rate limits, API pricing, or an explicit statement that no public access is offered yet.
Treat “Astra is GPT-6,” a precise Thursday launch time, leaked token rates, and claims that an endpoint has been “staged” as unconfirmed unless OpenAI publishes them. An API probe is not the same as a supported API product.
ChatGPT Astra FAQ
What is ChatGPT Astra?
ChatGPT Astra is the name OpenAI uses for an upcoming model with advanced long-horizon coding and cybersecurity capabilities. OpenAI has also linked an internal Astra version to ten reported mathematics and theoretical-computer-science advances.
Has ChatGPT Astra been released?
Not as a generally available ChatGPT or API product in the official material reviewed on September 3, 2026.
When will ChatGPT Astra launch?
OpenAI has not announced an exact release date. Online dates remain unconfirmed until OpenAI publishes access instructions or a live product/API update.
Is Astra GPT-6?
OpenAI has not officially identified Astra as GPT-6. “GPT-6” is a community label or prediction, not a confirmed public model name.
Can I use Astra in ChatGPT or through an API?
No publicly documented Astra model-picker entry, API endpoint, or model ID appears in the reviewed official sources. Selected testers and approved defensive cybersecurity users may receive controlled access, but that is not general availability.
How much will Astra cost?
OpenAI has not published Astra pricing. The reported $2,000 figure is an estimate for successful research-run tokens at GPT-5.6 Sol API rates, not a ChatGPT subscription fee or Astra API price.
Was Astra involved in the Hugging Face incident?
OpenAI’s August 7 cyber-capability update says Astra was not involved in exploiting Hugging Face. The company describes Astra’s safety work and the incident as part of the reason it strengthened controls, but those are separate claims.
What did Astra reportedly do in mathematics?
OpenAI reported ten advances spanning sphere packing, coding theory, group theory, quantum complexity, lattice problems, and extremal combinatorics. The accompanying Lean formalizations make the stated arguments machine-checkable, but they do not turn the research announcement into a public Astra trial.
If you need a model today, use a documented, available system rather than designing a workflow around a rumored Astra endpoint. GPT-5.6 Sol is an immediately testable baseline; Astra remains a watch item until OpenAI publishes its access terms, system card, and release path.