OpenAI's GPT-6 Astra Raises the Bar for AI Coding Agents

OpenAI used a September 3 press briefing to launch GPT-6 Astra, its newest flagship model, with president Greg Brockman telling reporters, "For me personally, I do think we're there... I think it's not unreasonable to feel that we are now in the AGI era." He was careful to frame that as a personal read, not a contractual or official company milestone. The headline framing aside, the benchmark numbers OpenAI published are the part worth sitting with if you build software with AI in the loop.

The coding and agent numbers

Astra scored 74.1% on DeepSWE v1.1, a software-engineering benchmark, and OpenAI says it completes software-engineering tasks at 57% lower cost per task than its predecessor. On OSWorld 2.0, a benchmark for agentic computer-use — operating a browser, spreadsheet, or desktop app the way a person would — Astra hit 72.6% accuracy in an average of 40 minutes per task, compared with the prior model's 65.7% in 75 minutes. That combination (faster *and* more accurate) is the more interesting signal than either number alone, since most agent upgrades until now have traded one for the other.

Brockman also pitched a reframe of how these tools should be priced: "What you actually want… is the price per task. [Token pricing] doesn't make any sense anymore." API access is priced at $10 per million input tokens and $50 per million output tokens in standard mode, with a $20/$100 'fast mode' tier, rolling out first to enterprise customers via OpenAI's Daybreak program before reaching ChatGPT and cloud-platform users.

The cybersecurity number that matters more

Astra is the first OpenAI model to cross the company's self-defined 'critical cybersecurity capability threshold' — its internal bar for a model that can find previously unknown vulnerabilities and chain them into working exploits across hardened systems without continuous human guidance. It scored 100% on ExploitBench, and OpenAI says testers discovered two previously unknown vulnerabilities during evaluation. Access to the model's full offensive-security capability is initially gated through a restricted program OpenAI calls Daybreak Blue, aimed at critical-infrastructure defenders. On the safety side, OpenAI reported 0% scope-exceeding incidents in internal testing for Astra, versus 48.2% for its immediate predecessor.

Why this matters for builders, not just OpenAI

For anyone doing AI-assisted development — what we call vibecoding — the practical read isn't "the model is smarter now." It's that the same capability curve driving better autocomplete and better agentic task completion is also driving better automated vulnerability discovery. Those are the same underlying skill: reading code, understanding intent, and generating a correct next step. A model that's meaningfully better at writing exploit chains is, by the same measure, better at spotting the bugs your own AI-generated code ships with — which is an argument for tightening review gates around AI output, not loosening them just because the tooling got faster.

The bottom line

Treat the 'AGI era' line as marketing color. Treat the DeepSWE, OSWorld, and ExploitBench numbers as the actual news: coding agents are getting both faster and more accurate at the same time, and the security implications of that are showing up in OpenAI's own safety disclosures, not just in speculation about them.