- By Prateek Levi
- Fri, 04 Sep 2026 12:55 PM (IST)
- Source:JND
GPT-6 Astra by OpenAI has now landed, and the company is claiming it to be its most capable AI suite to date. This, the company says, is a stepping stone towards its plan to reach AGI (artificial general intelligence), a stage where a computer gains human-like cognitive skills to reason and perform tasks. GPT-6 Astra is made keeping in mind that it will perform coding, reasoning, research and agentic workloads. When it comes to the token context window its stands at 1.05 million, and output tokens through the API stand at 128,000.
GPT-6 Astra: What's New?
The company, through a blog post, shared that this new model is capable of verifying security flaws that were previously unknown and also developing ways to manipulate them across multiple guarded systems by creating autonomous exploit chains like a real hacker.
ALSO READ: Airtel Offers Free Super Duolingo Subscription For 12 Months: Check Details
It can also do this with browsers by creating a browser compromise chain that can manoeuvre through the sandbox and was able to execute commands on the host. In another test, it was capable of creating a localised privilege-escalation chain. The GPT-6 Astra was able to score a 100 per cent score on ExploitBench, said the company, building upon an internal evaluation.
Additionally, this model is capable of performing high-level research and reasoning workloads. The API of this model is image- and text-compliant and also supports streaming, function calling and structured outputs.
This model is also touted to be the first to reach Critical level under its Preparedness Framework. But because of the model's capabilities, OpenAI has introduced multiple safeguards this time so it is not used by bad actors to cause problems and is more resistant to jail breaks that the GPT-5.6 Sol. The model refused 91.5 per cent of such requests in its cyber-jailbreak evaluations; in previous models, the same was 59 per cent.
The model has also reportedly shown around half as many higher-severity misalignment flags in a simulation involving more than 54,000 internal Codex tasks.
However, GPT-6 Astra still has some weak spots. The company says it was able to evade chain-of-thought monitoring in some adversarial tests, including situations where it deliberately underperformed or tried to carry out certain sabotage tasks. OpenAI says it has not seen any evidence of steganographic reasoning so far, but admits that GPT-6 Astra is not as easy to monitor as GPT-5.6 Sol.
ALSO READ: Mivi One 5G India Launch On September 24: Snapdragon Chip, Google Partnership Confirmed
GPT-6 Astra has a knowledge cutoff of April 30, 2026. The model is first being offered to enterprise customers through the Trusted Access Programme. API access, along with availability for Plus, Pro, Business and Enterprise users, is expected to roll out over the next few days.