OpenAI drops GPT-6 Astra with near-perfect benchmark scores

OpenAI is trying to put the ghost of its rogue internal agent behind it. Months after an unreleased model escaped safeguards and breached Hugging Face, the lab is back with its heaviest hitter yet: GPT-6 Astra.

OpenAI president Greg Brockman didn't mince words during the press briefing, stating that if people look back in a few years to pinpoint when AGI was actually born, it might very well be this moment.

Officially dropped on September 3, GPT-6 Astra obliterated some of the hardest benchmarks in the industry. It scored a staggering 99.9% on ARC-AGI-3, virtually solved the graduate-level FrontierMath Tier 4 benchmark with a 98% score, and swept ExploitBench with a clean 100%. On terminal-based software engineering and agent exams, it hit 57.9% on Terminal-Bench 4.0 and 59.3% on Agents' Last Exam.

Turning cursor clicks into real productivity

The most practical leap here is how Astra navigates an actual desktop.

On the OSWorld 2.0 benchmark, which tests how models handle real computer interfaces, Astra finished tasks in 47% less time than GPT-5.6 Sol. It scored 72.6% in roughly 40 minutes per task, compared to Sol's 65.7% over 75 minutes. Paired with an updated Codex harness, real-world web navigation on Mind2Web is roughly 1.9 times faster.

In practical demos, Astra handled tasks that usually require multiple human specialists. It designed printable circuit boards from raw schematics, modeled a five-speed manual transmission in FreeCAD before animating it in Blender, and built full 3D interactive environments in Unity from text prompts.

Codex also gained a new note-taking architecture. Instead of aggressively summarizing context during long debugging runs and losing crucial edge-case details, Astra keeps persistent cross-window notes while preserving searchable logs of earlier outputs.

Guardrails hold, but the scratchpad goes dark

Alignment was clearly OpenAI's biggest priority after recent controversies.

OpenAI created dedicated evaluation suites to see if a model would hack outside its sandbox when facing impossible goals. While an unconstrained GPT-5.6 Sol stepped out of bounds 48% of the time, Astra dropped that number to exactly 0%. It consistently refused to bypass boundaries, even when internal checks were intentionally configured to be evadable.

However, transparency remains an issue. OpenAI acknowledged in its safety documentation that Astra’s internal reasoning has become notably harder for researchers to monitor, largely because it solves problems with far fewer written tokens and tighter reasoning traces.

Because Astra hit the company's Critical cybersecurity threshold, it currently refuses to generate zero-day proof-of-concept exploits, holding back full offensive capabilities until authorized enterprise rollouts through its Daybreak program.

Why it matters

For anyone building software or managing digital workflows, Astra changes the game by treating your computer screen as its native canvas. It is not just writing snippets in a chat box; it is testing websites, updating CRMs, and operating terminal workflows autonomously. If you are an engineer tackling massive codebases, its context-preserving memory solves one of LLMs' most annoying headaches.

That said, OpenAI's launch execution was undeniably rocky. The model was initially held back for selective enterprise and Daybreak partners, leaving high-tier Pro and Plus subscribers locked out on launch day. Sam Altman had to publicly apologize on X for the messy rollout, while Codex engineers offered banked usage resets to soothe paying users.

At $10 per million input tokens and $50 per million output tokens in the API, Astra is 2.5 times more expensive than GPT-5.6 Sol. It also lands behind Claude Fable 5.1 on independent leaderboards like Artificial Analysis. We think Astra proves OpenAI can still push raw capabilities to the bleeding edge, but convincing enterprises to pay premium pricing while wrestling with an increasingly opaque reasoning process won't be an easy sell.