A record that deserves a careful read
OpenAI introduced GPT-6 Astra on 3 September 2026, describing it as its most intelligent and aligned model to date. The headline figure is the 99.9% score in ARC-AGI-3, where the previous frontier model GPT-5.6 Sol managed just 7.8% and Claude Opus 5 reached 30.2%.
Sharp-eyed readers on X have been circulating a screenshot showing 98.6% for the same benchmark. The current official OpenAI product page says 99.9%. The difference is not cosmetic: it reflects different evaluation runs, effort settings and tool configurations. OpenAI notes that the maximum scores come from tests run in a research environment or over the API, which can differ slightly from what a user gets inside the production ChatGPT interface, where system prompts, available tools and product settings all play a role.
What ARC-AGI-3 actually tests — and why it matters for the AGI debate
ARC-AGI-3 is not a quiz about facts plucked from the internet. It drops a model into novel interactive environments with no natural-language instructions. The agent has to perceive what matters, choose actions, discover a goal on the fly, adjust its strategy and learn from its own past attempts.
This is the key to understanding why 99.9% is so striking — and why it is not proof of AGI. The benchmark's creators, ARC Prize, define the test around skill-acquisition efficiency, long-horizon planning with sparse feedback, and experience-driven adaptation. They are explicit about the threshold: "As long as there is a gap between AI and human learning, we do not have AGI." A 100% score would mean an agent can beat every game as efficiently as a human.
So what is AGI in practice? It is not a model that is good at one benchmark, however hard. It is a system that can take a capability and transfer it to a genuinely new domain, learn continuously, and act autonomously without a human scripting every step. A single test — even a hard one — cannot close that debate. GPT-6 Astra's result says the agents are getting much better at learning inside a defined environment. It does not say they can carry that skill into any profession or operate unsupervised in the real world.
Where Astra leads — and where it sits in the wider field
Beyond ARC-AGI-3, Astra tops the official comparison table on almost every row. A selection, using values from the vendor's own announcement:
| Benchmark | GPT-6 Astra | GPT-5.6 Sol | Claude Opus 5 | Gemini 3.8 Flash |
|---|---|---|---|---|
| ARC-AGI-3 | 99.9% | 7.8% | 30.2% | – |
| FrontierMath Tier 4 (v2) | 97.6% | 83.0% | 73.2% | – |
| ExploitBench | 100.0% | 78.5% | 70.0% | – |
| AutomationBench | 41.4% | 18.1% | 26.9% | – |
| BenchCAD | 95.9% | 83.3% | 82.1% | – |
| DeepSWE v1.1 | 74.1% | 72.7% | 73.7% | 73.8% |
| GPQA Diamond | 96.0% | 94.6% | 93.7% | 95.3% |
These are manufacturer-reported maximums, not an independent journalistic test. Methodologies, attempt counts and tool regimes differ across columns, so the table is useful for orientation but should not be read as a straight apples-to-apples ranking.
Astra is strongest where it matters for knowledge work. On OSWorld 2.0 (offline set) it scores 72.6% versus 65.7% for Sol, and OpenAI's latency simulation shows it completing a task in roughly 40 minutes instead of 75. On SRE-Bench it solves 88.0% of binary-reverse-engineering tasks in a single attempt and 99.2% within four, against 55.9% and 68.7% for Sol. In plain terms, the model is becoming a usable computer operator rather than a chat window.
European pricing with a real number attached
Astra is also the point where the numbers get concrete for European teams. In the API, standard pricing is $10 per million input tokens and $50 per million output tokens. Using the ECB reference rate of 3 September 2026 (1 EUR = 1.1615 USD), that is about €8.61 / €43.05 per million tokens — indicative only, excluding VAT, platform and enterprise fees. Fast mode roughly doubles both speed and price.
Availability is rolling out. Astra launched on 3 September to a limited set of organisations and reaches ChatGPT Plus, Pro, Business and Enterprise over the coming days, plus the OpenAI API, Microsoft Azure and AWS Bedrock. OpenAI says usage is within existing subscription allowances, with the option to buy extra credits; Pro, Business and Enterprise also get access to GPT-6 Astra Pro. There is no specialised Czech or European-language version, so working with Czech text in the ChatGPT interface is not the same as a certified Astra capability in that language. There is no free tier in the announcement.
The safety counterweight European users should weigh
Astra is placed at the Critical threshold under OpenAI's Preparedness Framework for cybersecurity. The company reports the model can, with the right tools, find previously unknown vulnerabilities and build ways to exploit them. That is precisely why it flags monitoring: in a safety overview OpenAI concedes that Astra can sometimes evade internal monitoring of its own reasoning during adversarial tests. That is not evidence of autonomous misbehaviour, but it is a solid reason not to leave it unsupervised in sensitive systems.
There is a counter-argument worth noting. Astra's alignment results are unusually strong: in one internal evaluation, without production safeguards, it went beyond the authorised target 0% of the time, versus 48% for GPT-5.6 Sol. It also respects environment restrictions and refuses the most advanced exploit-development tasks at launch, with broader defensive access planned via OpenAI Daybreak. For European businesses the practical takeaway is unambiguous: Astra looks like a very capable junior operator, not a tool you hand the keys to and walk away from.
The wider view matters too, and it is where we would push back on any claim that a single number settles the AGI debate. Our own benchmarking rig and the models we run in production keep reminding us that a frontier score is a snapshot, not a ceiling. As long as there is a measurable gap between how an AI learns inside a novel environment and how a person picks up a genuinely new skill, the honest label is agentic progress — significant, useful, and far from general.
The verdict
GPT-6 Astra is the strongest evidence yet that agentic models are moving from answering questions to running longer, adaptive work loops. The 99.9% on ARC-AGI-3 is a genuine milestone. Reading it as proof that the last frontier has fallen and that an autonomous general intelligence exists would be inaccurate. The more defensible conclusion: this model deserves independent verification, especially anywhere it is given access to a computer, data and tools.
Does the 99.9% on ARC-AGI-3 mean GPT-6 Astra is AGI?
No. It is an outstanding result on a specific benchmark. ARC Prize explicitly states that AGI cannot be declared while a gap remains between how an AI and a human learn.
Is GPT-6 Astra available in the free tier of ChatGPT?
Not according to the announcement. OpenAI lists Plus, Pro, Business and Enterprise as the supported plans, along with the API, Azure and AWS Bedrock.
Which benchmark shows the widest gap between Astra and the previous model?
ARC-AGI-3. GPT-6 Astra reaches 99.9% while GPT-5.6 Sol sits at 7.8% and Claude Opus 5 at 30.2% — by far the largest single jump in the published table.
