OpenAI has officially launched GPT-6 Astra, its most capable model to date with 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and a perfect 100% on ExploitBench. This definitely isn’t a minor update from GPT-5.6 Sol. OpenAI is calling it a generational leap.
Astra is built on advances across pre-training, reinforcement learning, and alignment. It’s designed to handle what OpenAI describes as “professional work”, i.e not just answering questions, but taking on complex, multistep tasks across software engineering, scientific research, computer use, and cybersecurity. The idea is that you delegate something meaningful to it and trust the outcome, which is a much harder bar to clear than most models have managed so far.
The computer use story is where Astra makes its most concrete gains. In latency simulations on OSWorld 2.0, Astra scored 72.6% at roughly 40 minutes per task, compared to GPT-5.6 Sol’s 65.7% at around 75 minutes. It’s a fundamentally faster model that can fill out forms, update CRM records, run frontend QA checks on websites, analyse scientific data, and handle calendar management without needing constant hand-holding. Combined with an updated Codex harness, task completion on the Mind2Web benchmark is reportedly 1.9x faster than the current GPT-5.6 Sol experience.
For software engineering, Astra outperforms GPT-5.6 Sol on DeepSWE v1.1 at approximately 57% lower estimated API cost per task. One of the more interesting technical changes is a new memory system in Codex: rather than relying on compaction (which can lose important context during long sessions), Astra can now keep notes across context windows and search earlier context for relevant details.
Canva, which was among the first companies to get early access, put Astra through a serious test. Danny Wu, Canva’s Head of AI Products, described it navigating a codebase of over 80 million lines of code, running more than a thousand queries against their data warehouse, and drawing on 21 different internal knowledge sources to produce what he called some of the best and most signal-rich recommendations they had seen.
On that front, Astra is described as OpenAI’s most aligned model yet. In an evaluation inspired by a real-world incident involving Hugging Face, GPT-5.6 Sol (without production safeguards) exceeded its authorised scope 48% of the time. Astra did so 0% of the time. It’s also three times less likely than GPT-5.6 Sol to make inaccurate representations about its own capabilities. In Codex, it can ask clarifying questions asynchronously while continuing work that doesn’t depend on the answer, and it will hold on consequential decisions rather than proceeding with guesswork.
The cybersecurity picture is the most complex part of this launch. Astra meets the Critical threshold under OpenAI’s Preparedness Framework, which means its offensive capabilities are serious enough to require extra care. Without production safeguards, it scored 100% on ExploitBench and even discovered two previously unknown zero-day vulnerabilities during evaluation, which OpenAI says are being disclosed to the relevant maintainers. With the version launching today, defenders can use it for secure code review and patching. More advanced workflows, including vulnerability validation, malware analysis, and detection engineering, will roll out via OpenAI Daybreak in the coming weeks as safeguards are updated to match.
For scientific work, Astra scored 98% on FrontierMath Tier 4 and has already contributed to new mathematical knowledge, including two further results on prime number gaps being shared today. It can also work directly in specialised scientific software, combining reasoning with hands-on computer use to help researchers explore data and decide what to investigate next.
GPT-6 Astra is rolling out now to a limited set of organisations, with broader availability coming to ChatGPT Plus, Pro, Business, and Enterprise users over the coming days. It will also be available through the OpenAI API, Microsoft Azure, and AWS Bedrock. API pricing is set at USD $10 per million input tokens and USD $50 per million output tokens. A Fast mode is also available, delivering up to twice the processing speed at twice the Standard price. No Australian dollar pricing has been announced, but existing subscription allowances cover Astra usage, with the option to purchase additional credits. Pro, Business, and Enterprise users will also get access to GPT-6 Astra Pro. Enterprise admins should note that access is off by default at launch and needs to be enabled manually.
Read more on the official blog.
