GPT-6 Astra: What Is OpenAI's New Model Really Worth?

On 3 September 2026, OpenAI presented GPT-6 Astra, which it describes as the world's most intelligent and aligned model. President Greg Brockman closed the press conference with a striking phrase: “Welcome to the AGI era.” But the published scores vary by more than a factor of two depending on the test configuration. Here is what the verified numbers say, and what this model really changes.

---

OpenAI's largest training campaign

GPT-6 Astra was trained on more than 100,000 GPUs at the Stargate facility in Texas, the largest computing infrastructure OpenAI has ever used, according to Axios and Ina Fried on 3 September 2026. The model has a 1,050,000-token context window and can produce up to 128,000 output tokens. Its knowledge cutoff is 30 April 2026.

Astra is available in two versions:

• GPT-6 Astra (standard): available through ChatGPT Plus, Pro, Business, Enterprise, and the API.
• GPT-6 Astra Pro: reserved for Pro, Business, and Enterprise subscribers.

At launch, Astra was first rolled out to a limited group of organizations through the Daybreak Access program, before gradually opening to all subscribers over the following days, according to Axios.

Record scores, under specific conditions

OpenAI published a series of benchmarks in which Astra significantly outperforms GPT-5.6 Sol, released in late July 2026, according to Engadget. But the scores were not all obtained under the same conditions. Here are the verified results:

Benchmark	GPT-6 Astra	GPT-5.6 Sol	What it measures
ARC-AGI-3 (Standard harness)	62.7%	7.8%	Interactive agentic reasoning
ARC-AGI-3 (Provider Adapter)	99.9%	38.3%*	Same test with optimized context handling
FrontierMath Tier 4	98%	Not disclosed	Open mathematical problems
Terminal-Bench 4.0	57.9%	37.3%	Agentic terminal tasks
Agents' Last Exam	59.3%	53.6%	General agentic skills
ExploitBench	100%**	5.5%	Offensive cybersecurity

Sources: OpenAI announcement, 3 September 2026; ARC Prize Foundation, 3 September 2026; BenchLM.

• Score on the public set with OpenAI's Responses API, using reasoning retention and compaction. OpenAI published it on 29 July 2026. It is not directly comparable with Astra's Semi-Private score.

** OpenAI notes that the “100% success rate may not be achievable” because some vulnerabilities in the benchmark do not allow code execution within the evaluation constraints, as stated in note 14 of the announcement.

Important: the ARC-AGI-3 figures cannot be compared line by line. The 99.9% score was obtained with a Provider Adapter harness that preserves the model's internal reasoning state between actions, while the Standard harness resets it on every turn. The next section explains why that difference matters.

Two harnesses, two scores: why the difference matters

The ARC Prize Foundation, which created the ARC-AGI-3 benchmark, published a detailed analysis of Astra's results. Two configurations were tested:

1. Standard harness: the model can carry text notes between actions, but its private reasoning is reset after every turn. Score: 62.7%, at an estimated cost of $26,098.
2. Provider Adapter harness: the model preserves its opaque reasoning state between requests and uses compaction to manage long histories. Score: 99.9%, at an estimated cost of $18,817.

With the Provider Adapter, Astra even exceeds human action efficiency: it uses fewer actions than the human median on 96% of levels, and 51.7% fewer actions on average, according to the ARC Prize Foundation.

Greg Kamradt of the ARC Prize Foundation said: “Astra surpassed our human action-efficiency baseline on 96% of levels, reaching human parity on the benchmark.”

This difference illustrates a key point: benchmarks measure not only the model, but also the infrastructure around it. A model that performs exceptionally in an optimized harness can look mediocre in a generic harness.

Pricing and availability

GPT-6 Astra is available through:

• The OpenAI API: $10 per million input tokens and $50 per million output tokens in standard mode. Fast mode, which is 2.5 times faster, doubles those prices.
• ChatGPT: included in existing subscriptions.
• Microsoft Azure and AWS Bedrock: available to enterprise customers.

OpenAI argues that the cost per completed task matters more than the cost per token. On DeepSWE v1.1, Astra cost approximately 57% less per software engineering task than Sol, according to the OpenAI announcement.

Safety: the first model at the Critical level

GPT-6 Astra is the first OpenAI model to reach the Critical threshold of the Preparedness Framework for cybersecurity. This means that, with the right tools, Astra can find unknown vulnerabilities in well-protected systems and develop new exploitation methods without step-by-step human guidance, according to OpenAI's Safety Overview.

OpenAI has strengthened its safeguards:

• Strict isolation and encryption of checkpoints
• Universal monitoring of complete trajectories, including reasoning chains
• Blocking alignment evaluations before internal use
• An external monitoring system to detect misaligned behavior

The model is also more resistant to jailbreaks than GPT-5.6 Sol. In a simulation of 54,000 Codex tasks, Astra received approximately half as many reports of misaligned behavior as Sol, according to the System Card.

There is still a concern: OpenAI acknowledges that Astra is less monitorable than previous versions, meaning that it is harder to monitor when it hides its intentions. Improving monitorability remains a research priority.

“Welcome to the AGI era”

The most widely discussed statement came from Greg Brockman: “I think this model could mark the arrival of AGI. Welcome to the AGI era.”, according to Axios.

OpenAI defines AGI as a system capable of performing most economically useful tasks at least as well as a human. Astra checks several boxes: it exceeds humans in action efficiency on ARC-AGI-3, solves open mathematical problems with a 98% FrontierMath score, and achieves a perfect score in offensive cybersecurity with 100% on ExploitBench.

But OpenAI itself acknowledges the limits. Brockman said there is no clearly defined “AGI moment”. The ARC Prize Foundation added: “We have made it clear that saturating the benchmark would not constitute proof of AGI. Astra represents significant progress toward generalization, but we are not claiming that it is AGI.”

What changes for businesses?

1. Automating complex tasks becomes realistic. With 57.9% on Terminal-Bench 4.0 and demonstrated capabilities in web navigation and tool use, Astra can handle complete workflows.
2. The cost-efficiency ratio improves. Despite a higher per-token price than Sol, the cost per completed task is lower.
3. Security becomes central. The Critical cyber level forces companies to think seriously about isolation and agent monitoring.
4. Evaluation infrastructure matters as much as the model. The score gap, 62.7% versus 99.9%, shows that the harness, context management, and API parameters are levers just as important as the model itself.

Sources and methodological notes

This article was written from the following sources, all checked on 11 September 2026:

• OpenAI announcement: https://openai.com/index/gpt-6-astra/
• Safety report: https://openai.com/index/safety-overview-gpt-6-astra/
• Enterprise overview: https://openai.com/index/gpt-6-astra-next-generation-work
• System Card: https://deploymentsafety.openai.com/gpt-6-astra
• ARC Prize analysis: https://arcprize.org/blog/astra
• Axios, Ina Fried: https://www.axios.com/2026/09/03/openai-astra-gpt-6-agi-brockman
• The Decoder, Matthias Bastian: https://the-decoder.com/gpt-6-astra-is-the-first-model-making-openai-willing-to-declare-the-agi-era
• API documentation: https://developers.openai.com/api/docs/models/gpt-6-astra

Verified facts: benchmark scores, pricing, availability dates, and Greg Brockman's public statements come from primary sources, including OpenAI and the ARC Prize Foundation, or verified secondary sources such as Axios, The Decoder, and Engadget.

Editorial analysis: the interpretation of the scores, comparison of the harnesses, the AGI debate, and the implications for businesses.

Limitations: ARC-AGI-3 scores obtained with the Provider Adapter harness are not directly comparable with Standard harness scores. The estimated per-task costs come from OpenAI and were not independently verified. ExploitBench's 100% is a benchmark ceiling, not necessarily absolute perfection, as OpenAI explains in its footnote.