What Is GPT-6 Astra?
GPT-6 Astra is OpenAI’s newest frontier large language model, unveiled and released as a limited preview on September 3, 2026. OpenAI calls it a “generational leap” across cybersecurity, professional work, software engineering, and science. The company plans a broader public release for paid users around September 7, 2026, alongside several related models.
Unlike a single model, the release ships as a family of six reasoning-level variants (max, xhigh, high, medium, low, and a Non-reasoning tier), each trading intelligence for speed and cost. Backed by a 1,000,000-token (1M) context window and OpenAI’s largest-ever training run on more than 100,000 GPUs at its Stargate site in Texas, it is designed for long, multi-step professional workflows rather than simple chat.
Key Specs at a Glance
These are the verified technical specifications reported by OpenAI and independent benchmarkers at Artificial Analysis on release day.
| Specification | Value |
|---|---|
| Developer | OpenAI (ChatGPT maker) |
| Release date | September 3, 2026 (limited preview) |
| Public rollout | Around September 7, 2026 (paid users) |
| Context window | 1,000,000 tokens (~1,500 A4 pages) |
| Knowledge cutoff | April 30, 2026 |
| Input modality | Text and image |
| Output modality | Text |
| Variant count | 6 (max, xhigh, high, medium, low, Non-reasoning) |
| Top reasoning variant | GPT-6 Astra (max) |
The Six GPT-6 Astra Variants
Artificial Analysis assigned each variant an Intelligence Index score (higher is smarter) and a price per standard Intelligence Index task. The scores show a tight, purposeful spread rather than one dominant model.
| Variant | Intelligence Index | Cost per task |
|---|---|---|
| GPT-6 Astra (max) | 61 | $1.67 |
| GPT-6 Astra (xhigh) | 61 | $1.20 |
| GPT-6 Astra (high) | 60 | $0.96 |
| GPT-6 Astra (medium) | 59 | $0.75 |
| GPT-6 Astra (low) | 57 | $0.46 |
| GPT-6 Astra (Non-reasoning) | 55 | $0.93 |
Prices vary by up to 3.6x across the family. Notably, the Non-reasoning tier sits between low and high on cost while delivering the lowest intelligence score, which makes it a budget-speed option for high-volume, non-reasoning tasks where latency matters more than deep reasoning.
How the Intelligence Index works
Artificial Analysis Intelligence Index v4.1.1 combines 9 evaluations: GDPval-AA v2, Tau-Banking, Terminal-Bench v2.1, SciCode, Humanity’s Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR. It is a broad measure of real-world capability, not a single leaderboard run.
A Generational Leap in Reasoning
OpenAI frames GPT-6 Astra as a step-change in capability. The company said it “is faster and capable of performing more tasks than any prior iteration,” and better at staying focused, adhering to task boundaries, understanding user intent, handling tedious work, and completing multi-step workflows.
Professional and Scientific Work
The biggest claimed gains are in professional environments. GPT-6 Astra pairs advances in computer use with targeted training for workplace tasks, meaning it can interact with the tools and systems professionals already use rather than only answering questions in a chat box. OpenAI’s president Greg Brockman went further, saying the model could eventually be seen as the arrival of artificial general intelligence (AGI) as OpenAI once defined it: “an automated system that can perform all economically valuable work as well as or better than humans.”
Software Engineering
With a 1M context window that fits roughly 1,500 pages and dual reasoning paths, the model is positioned for large codebases, long debugging sessions, and complex multi-file refactors. The larger context lets a single session hold an entire codebase context without aggressive summarization.
Cybersecurity Comes With New Guardrails
Cybersecurity is one area where OpenAI is deliberately holding back. On September 1, 2026, OpenAI reported that access to its most advanced cybersecurity capabilities will be more limited. Advanced offensive security work is initially restricted to a small group of testers, with access through its Daybreak Blue program to follow and expand defensive use cases.
Safety focus after the July incident
GPT-6 Astra’s launch was delayed from an earlier target after OpenAI’s Hugging Face incident in July 2026, so the team could add more safeguards. Those extra guardrails are part of why advanced capabilities ship in stages rather than all at once.
The Recurrent Depth Technique and Its Controversy
GPT-6 Astra introduces a new reasoning method called recurrent depth. According to OpenAI, it works in a way that obscures some or all of the AI’s reasoning, otherwise known as its “chain of thought.” Even in limited use, this has drawn concern from researchers about how monitorable (or auditable) the model is, since users and evaluators cannot easily inspect the step-by-step logic behind answers.
This is a trade-off at the frontier: opacity can improve performance and reject manipulation, but it makes safety auditing harder. It is worth watching how OpenAI and outside evaluators respond to calls for transparency as the model rolls out more widely.
An Unprecedented Training Run
Aidan Clark, OpenAI’s vice president of research, told reporters that GPT-6 Astra’s development involved “by far” their largest training run. It is the first time the company has pretrained on more than 100,000 GPUs at its Stargate site in Texas.
That scale is a signal. Frontier-model training runs are now measured in hundreds of thousands of accelerators, which has major implications for the compute, energy, and capital demands of the AI industry and for which organizations can realistically compete at this level.
How GPT-6 Astra Compares to Its Predecessors
GPT-6 Astra replaces and extends the GPT-5 generation. On Artificial Analysis, the top GPT-6 Astra variants score 61 on the Intelligence Index, matching the strongest GPT-5.6 release, GPT-5.6 Sol (also 61), while newer friends like GPT-5.6 Terra (57) and GPT-5.6 Luna (52) sit lower. The step up is more about the 1M context, multimodal input, and the reasoning-level pricing ladder than a dramatic raw-score jump at the top.
Model Top Intelligence Index GPT-6 Astra (max / xhigh) 61 GPT-5.6 Sol 61 GPT-6 Astra (high) 60 GPT-6 Astra (medium) 59 GPT-5.6 Terra / GPT-6 Astra (low) 57 GPT-6 Astra (Non-reasoning) 55 GPT-5.6 Luna 52 Astra vs GPT-5.6 Sol: What Changed?
Because both reach an Intelligence Index of 61, the choice between Astra and Sol comes down to use case rather than raw capability.
- Context: Both carry a 1M-token window, so long-document work is covered on either.
- Reasoning tiers: Astra offers six price-and-speed tiers, giving you more granular control over cost for low-stakes tasks (the $0.46 low tier) versus top reasoning (the $1.67 max tier).
- Computer use: Astra is explicitly trained to handle professional computer-use workflows, which is its headline advantage over earlier models.
- Safety posture: Astra ships with staged cybersecurity access and the recurrent-depth reasoning changes, marking a more careful release process than prior generations.
Who Should Use GPT-6 Astra?
The model is aimed primarily at professional and technical users, not casual chatbot interaction.
Good fits
- Software engineers working across large codebases who can exploit the 1M-token context.
- Researchers and scientists running complex, multi-step analytical tasks.
- Enterprise teams building computer-use agents that operate existing workplace software.
- Budget-conscious API users who can route easy tasks to the low or Non-reasoning tiers to cut cost per request.
Less ideal
- Simple Q&A or content generation where a cheaper, faster model is more economical.
- Use cases that require fully transparent, auditable reasoning given the recurrent-depth design.
- Anything needing the most sensitive cybersecurity automation right now, given the staged access.
One practical tip
If you are an API developer, start with the
loworhightier for most traffic and reservemaxfor the hardest reasoning tasks. The 3.6x price spread means routing your requests to the right tier can cut your bill dramatically without hurting quality on easy jobs.Access, Pricing, and Availability
GPT-6 Astra launched as a limited preview for trusted partners on September 3, 2026. OpenAI plans to open access to paid ChatGPT users around September 7, 2026, alongside the related family of models. API access is expected via the standard OpenAI platform, priced per reasoning tier as shown above.
Because availability is rolling out in stages and advanced capabilities are gated, expect differences in what individual accounts can access during the first days after launch. The fastest way to know if your tier is enabled is to check the model selector in ChatGPT or the models list in the OpenAI API.
Final Verdict
GPT-6 Astra is a genuinely significant release: the first OpenAI model pretrained on more than 100,000 GPUs, with a 1M context window, six reasoning-price tiers, and a headline focus on professional computer use. The trade-offs are real, from a more opaque reasoning technique to deliberately gated cybersecurity powers. For developers and teams doing complex, long-horizon work, it sets a new ceiling. For casual use, it is likely overkill, and you will be better off on a lower tier or a cheaper model.
Keep an eye on the September 7 public rollout and on how the community reacts to the recurrent-depth transparency question. That is where the next real story is likely to come from.
Frequently Asked Questions
GPT-6 Astra was released on September 3, 2026 as a limited preview for trusted partners. A broader public rollout for paid ChatGPT users is planned around September 7, 2026.
The release includes six models: max, xhigh, high, medium, low, and a Non-reasoning tier. Each trades intelligence for a different speed and price, from $0.46 (low) to $1.67 (max) per standard Intelligence Index task.
GPT-6 Astra supports a 1 million token context window, roughly equivalent to 1,500 A4 pages of text at size 12 Arial. Its knowledge cutoff is April 30, 2026.
Recurrent depth is a new reasoning technique OpenAI introduced with GPT-6 Astra. It obscures some or all of the model’s chain of thought, which has raised questions from researchers about how monitorable and auditable the model is.
It was OpenAI’s largest training run ever, the first time the company pretrained on more than 100,000 GPUs, using its Stargate site in Texas. That scale is a major signal for compute and energy demands in the AI industry.
OpenAI is deliberately limiting access to the most advanced cybersecurity capabilities. Advanced offensive security work is initially restricted to a small group of testers, with defensive uses expanding later through its Daybreak Blue program.
API access is expected on the OpenAI platform as the rollout proceeds, priced per reasoning tier. OpenAI plans to open access to paid ChatGPT users around September 7, 2026, alongside the related family of models.