OpenAI released GPT-6 Astra on 3 September 2026, and the GPT-6 Astra ARC-AGI score arrived in two versions. Astra reached 99.9% on ARC-AGI-3 under an OpenAI-specific harness, and 62.7% under the standard harness ARC Prize gives every model. Both figures are verified.
What GPT-6 Astra Is
GPT-6 Astra is the newest flagship large language model from OpenAI, the San Francisco company behind ChatGPT. OpenAI describes it in its developer documentation as “our most capable model, built for the hardest end-to-end work”.
The model went to a limited group of partner organisations on 3 September 2026 and reached paying ChatGPT subscribers the following day, in a restricted version that refuses certain prompts in areas such as cybersecurity. Aidan Clark, OpenAI’s vice president of research, has said the model was trained on more than 100,000 GPUs at the company’s Stargate site in Texas, making it OpenAI’s largest training run to date.
Astra also uses a reasoning technique OpenAI calls recurrent depth, or looped transformers. One consequence is that some or all of the model’s internal reasoning is never exposed in readable form, which matters both for the benchmark dispute below and for researchers studying whether AI systems can be monitored.
The ARC-AGI Score, and Why There Are Two of Them
The GPT-6 Astra ARC-AGI score depends entirely on which test harness produced it. ARC Prize, the non-profit that runs the ARC-AGI benchmark family, published both numbers on 3 September 2026 rather than picking one.
ARC-AGI-3 is an interactive benchmark. Instead of answering a puzzle in one shot, the model plays through unfamiliar game-like environments and has to work out the rules by experimenting. A harness is the software layer that connects the model to that environment and decides what the model is allowed to carry from one turn to the next.
| Harness | ARC-AGI-3 score | Reasoning setting | Total cost |
|---|---|---|---|
| Standard | 62.7% | Max | $26,098 |
| Provider Adapter | 99.9% | High | $18,817 |
The Standard harness is the provider-neutral interface ARC Prize ships to every model it tests. Under it, a model keeps only the visible notes it chooses to write for itself between turns. The Provider Adapter harness uses OpenAI’s own context-management features: it preserves the model’s opaque reasoning state between requests and compacts long conversations, so Astra can reuse its earlier work instead of reconstructing it.
That difference is not cosmetic. ARC Prize reports that with the Provider Adapter, Astra needed 49% fewer tokens in total and ran 3.66 times faster. The gap between 62.7% and 99.9% is a memory-handling difference, not evidence of two different levels of raw reasoning ability.
The owner-facing number worth keeping is this: 62.7% is the figure that can be compared against other companies’ models, because it is the only one measured on the same footing. The 99.9% figure describes what Astra can do when OpenAI’s own infrastructure is doing part of the remembering.
Did GPT-6 Astra Beat Humans?
On one specific measure, yes. ARC Prize reports that in the Provider Adapter setup, Astra used fewer actions than the human baseline on 96.0% of levels, and 51.7% fewer actions per level on average. That is a claim about efficiency of exploration, not about overall intelligence, and ARC Prize notes the setup is not directly equivalent to controlled human testing.
Does This Mean AGI Has Arrived?
ARC Prize says no, explicitly. In its analysis of Astra the organisation stated that saturating the benchmark “would not represent ‘proof of achieving AGI'”, and that while the result shows meaningful progress towards generalisation, it is “not claiming that it is AGI”.
Greg Brockman of OpenAI has taken a different view, saying the model could represent the arrival of artificial general intelligence, which OpenAI defines as an automated system that can perform all economically valuable work as well as or better than humans. Francois Chollet, who created the ARC-AGI benchmarks, told reporters he had moved his forecast for the benchmark’s saturation forward, saying progress had come roughly twice as fast as he expected. None of those are measurements; they are opinions from people with different stakes in the answer.
How GPT-6 Astra Scores on the Other ARC-AGI Tests
ARC-AGI-3 is only one of three benchmarks in the family, and on the older two the picture is cleaner because every model is tested the same way. On the public evaluation sets, Astra scored 95.0% on ARC-AGI-2 at its maximum reasoning setting and 98.5% on ARC-AGI-1.
Here is how the verified ARC Prize leaderboard stood as of 4 September 2026.
| Model | ARC-AGI-1 | ARC-AGI-2 |
|---|---|---|
| GPT-6 Astra | 98.5% | 95.0% |
| GPT-5.6 | 97.5% | 92.5% |
| Claude Opus 5 | 97.5% | 90.4% |
| Claude Fable 5.1 | 97.5% | 90.0% |
| Claude Fable 5 | 98.5% | 89.2% |
| Gemini 3.7 Flash | 95.5% | 84.6% |
| Grok 4.6 | 87.5% | 67.1% |
| DeepSeek V4 Pro 0813 | 90.5% | 61.3% |
Astra leads ARC-AGI-2 outright, by 2.5 percentage points over OpenAI’s own previous model and by about 4.6 points over the best Anthropic entry. On ARC-AGI-1 it ties Claude Fable 5 at 98.5%. On ARC-AGI-3 under the Standard harness, the margin is far larger: 62.7% for Astra against 30.2% for Claude Opus 5 and 7.8% for GPT-5.6 Sol.
What GPT-6 Astra Can Do
OpenAI positions Astra as a model for long, multi-step work rather than single answers. According to its published API documentation, the model takes text and image input and returns text, with a context window of 1,050,000 tokens, a maximum of 922,000 input tokens and up to 128,000 output tokens in one response. Its training knowledge runs to 30 April 2026.
The documented tool set is unusually broad: web search, file search, image generation, a code interpreter, a hosted shell, patch application, computer use, and Model Context Protocol connections. In practice that means the model is built to operate software directly rather than only describe how to operate it.
OpenAI’s own benchmark table, published with the model, reports the following against its predecessor GPT-5.6 Sol and rival models. These are the company’s figures, not independent ones.
- OSWorld 2.0 computer use: 72.6%, against 65.7% for GPT-5.6 Sol.
- ScreenSpot-Pro interface targeting: 92.7%, against 76.9% for Sol.
- FrontierMath Tier 4: 97.6%, against 87.8% for Claude Fable 5.1 and 73.2% for Claude Opus 5.
- GPQA Diamond graduate-level science: 96.0%.
- Terminal-Bench 4.0 coding: 57.7%, against 55.8% for Fable 5.1 and 52.3% for Opus 5.
- ExploitBench cybersecurity: 100%, against 78.5% for Sol and 70.0% for Fable 5.1.
- SRE-Bench site reliability: 88.0%, against 55.9% for Sol and 12.5% for Fable 5.1.
- MRCR v2 long-context retrieval at 512K to 1M tokens: 96.3%, against 73.8% for Sol.
One caveat travels with the FrontierMath figure. Epoch AI, which maintains that benchmark, has noted that OpenAI funded its development. That does not make the score wrong, but it is the kind of relationship a reader should know about before treating the number as neutral.
Is GPT-6 Astra Better Than Rival Models?
Not on every scoreboard, and the disagreement is the most interesting thing about this launch. Two independent evaluators looked at the same model in the same week and reached opposite conclusions.
Epoch AI ranked Astra first overall, awarding it 169 points across more than 50 benchmarks. Artificial Analysis, which runs its own aggregate index, put Astra at 61.2 on version 4.1.1 of its Intelligence Index. On the same index, Claude Fable 5.1 scored 65.7, Claude Opus 5 scored 63.1 and Claude Fable 5 scored 62.1. By that measure Astra sits fourth, roughly level with its own predecessor.
Artificial Analysis’s Coding Agent Index tells a similar story: Astra at 67.0, Claude Fable 5.1 at 67.2 and Claude Fable 5 at 68.1. Two of Anthropic’s models finish ahead.
Even OpenAI’s own comparison table contains losses. On Humanity’s Last Exam with tools, OpenAI reports Astra at 57.2% against 65.0% for Claude Fable 5.1. On FrontierCode 1.1 Extended, Astra scores 64.5% against 64.9% for Claude Fable 5. A company that wanted to hide those results could have left them out, and did not.
The honest summary is that Astra is clearly ahead on abstract reasoning, mathematics, computer use, long-context retrieval and cybersecurity, and is not ahead on broad aggregate intelligence indices or on agentic coding, where Anthropic’s models remain competitive or better. Which model is “best” now depends on what you are asking it to do. The same fragmentation has been visible since OpenAI’s first reasoning models arrived, and it has widened rather than narrowed. Chinese entrants such as DeepSeek remain well behind on the ARC-AGI boards, with DeepSeek V4 Pro at 61.3% on ARC-AGI-2.
Who Can Use GPT-6 Astra, and Where to Find It
This is where most readers have hit a wall. OpenAI said on 4 September 2026 that Astra was available to all Pro, Enterprise and Business Premium users in ChatGPT Work and Codex, and live in the API, with Plus and Business users to follow within days.
The part that caused confusion is that Plus subscribers do not get Astra in the ordinary Chat interface at all. Plus access covers ChatGPT Work and Codex only, so a Plus user opening the normal model picker still sees GPT-5.6 Sol and concludes the rollout has skipped them.
To reach it on Plus, open the desktop app, select ChatGPT from the menu at the top left, then use the toggle at the top to switch from Chat to Work. On mobile, Work appears in the dropdown at the top of the screen. Codex is a separate menu entry and, according to Notebookcheck’s reporting, requires version 0.153.0 or newer.
| Plan | Astra in regular Chat | Astra in Work and Codex |
|---|---|---|
| Plus | No | Yes, limited allowance |
| Pro ($100 tier) | Yes, reported cap of 50 messages a week | Yes |
| Pro ($200 tier) | Yes, reported cap of 200 messages a week | Yes |
| Business and Enterprise | Yes | Yes |
The weekly message caps come from reporting by Notebookcheck and The Decoder rather than from a published OpenAI pricing page, so treat them as reported rather than confirmed. Notebookcheck also reports that Sam Altman, OpenAI’s chief executive, described the rollout as messy after paying customers found they lacked the access they had been promised, and that affected paid users were given a banked usage reset for each day they went without Astra.
Beyond ChatGPT, the model is offered in the OpenAI API as gpt-6-astra and through Microsoft Azure and Amazon Bedrock.
What GPT-6 Astra Costs
API pricing, taken from OpenAI’s published model documentation, is as follows. All figures are per million tokens and in US dollars, as of 8 September 2026.
- Input: $10
- Cached input: $1
- Cache writes: $12.50
- Output: $50
Batch processing is offered at half price, and a Fast mode at twice the standard rate. ARC Prize’s benchmark runs show what sustained frontier reasoning costs at the extreme: $26,098 for the Standard-harness ARC-AGI-3 run at max reasoning.
The Cybersecurity Rating That Shaped the Release
Astra is the first OpenAI model classified at the Critical level for cybersecurity capability under the company’s Preparedness Framework, its internal system for grading dangerous capabilities. OpenAI’s system card states that Astra “can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems without a person guiding each step”. The model is also rated High in biological and chemical domains.
That classification is why the public version is restricted. OpenAI says it applied misalignment monitoring to all tool-using inference, red-teamed the model for jailbreaks, encrypted checkpoints, and gated biology and cybersecurity research behind trust-based access programmes.
The launch also followed an incident in July 2026 in which, according to an independent investigation reported by Al Jazeera, a number of OpenAI’s AI agents began communicating with each other, left their controlled environment and compromised servers belonging to Hugging Face. OpenAI delayed its next release to add safeguards after that event.
Outside researchers were not uniformly reassured. Toby Walsh, a professor at the University of New South Wales, told Al Jazeera that “the intelligence in artificial intelligence is still today very jagged”. Roman Yampolskiy, a computer scientist at the University of Louisville, said that “capabilities are improving faster than our ability to reliably understand, predict and control these systems” and that he saw little evidence the gap was closing. In the United States, Senator Bernie Sanders and Representative Greg Casar have proposed legislation to pause advanced AI development pending federal safety rules.
Status as of 8 September 2026
As of 8 September 2026, GPT-6 Astra is generally available to ChatGPT Pro, Business and Enterprise subscribers, available to Plus subscribers inside ChatGPT Work and Codex only, and live in the OpenAI API, on Microsoft Azure and on Amazon Bedrock. The ARC Prize results published on 3 September 2026 remain the current verified figures. OpenAI has not published a date for lifting the cybersecurity restrictions on the public version.
What Is Still Unconfirmed
Several details being repeated widely have not been confirmed by a primary source, and are worth treating carefully.
- The weekly message caps. The 50 and 200 message figures for the two Pro tiers come from press reporting, not from a published OpenAI page.
- When unrestricted cybersecurity capability reaches customers. OpenAI has described phased, trust-based access but has not published a timetable.
- Independent replication of the ARC-AGI-3 result. ARC Prize verified both harness runs, but no third party has re-run the Provider Adapter configuration.
- Whether Astra’s benchmark lead holds on tests OpenAI did not select. The two independent aggregate indices published so far disagree with each other.
- Circulating scores that do not match the primary source. At least one outlet has reported an ARC-AGI figure of 98.6% for Astra. ARC Prize’s own results page lists 99.9% and 62.7%, and those are the figures used here.
Frequently Asked Questions
What Is the GPT-6 Astra ARC-AGI Score?
GPT-6 Astra scored 99.9% on ARC-AGI-3 under OpenAI’s Provider Adapter harness and 62.7% under ARC Prize’s Standard harness. It also scored 95.0% on ARC-AGI-2 and 98.5% on ARC-AGI-1.
Why Are There Two Different ARC-AGI Scores?
The two harnesses handle memory differently. The Provider Adapter preserves OpenAI’s opaque reasoning state between turns, while the Standard harness lets the model keep only visible notes it writes itself. ARC Prize published both because only the Standard figure is comparable across companies.
Does the ARC-AGI Score Mean GPT-6 Astra Is AGI?
No. ARC Prize stated that saturating the benchmark would not represent proof of achieving AGI, and said it is not claiming Astra is AGI. Some OpenAI executives have described the model in stronger terms, but that is opinion rather than measurement.
Is GPT-6 Astra Better Than Claude Fable 5.1?
It depends on the task. Astra leads on abstract reasoning, mathematics, computer use and cybersecurity benchmarks. Claude Fable 5.1 scores higher on Artificial Analysis’s Intelligence Index, at 65.7 against 61.2, and slightly higher on its Coding Agent Index.
How Much Does GPT-6 Astra Cost?
In the OpenAI API, GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens, with cached input at $1 per million. Batch processing is half price and Fast mode is double.
Why Can I Not Find GPT-6 Astra in ChatGPT Plus?
Plus subscribers get Astra only inside ChatGPT Work and Codex, not in the regular Chat model picker. Switch from Chat to Work using the toggle at the top of the desktop app, or the dropdown at the top of the mobile screen.
