GPT-6 Astra ships with computer-use — but it still waits to be told what to do

OpenAI released GPT-6 Astra on September 3, 2026, positioning computer-use as the headline capability: the model fills forms, navigates web pages, updates CRM records, and drafts email through the Responses API harness. The number that anchors the launch is 72.6% on the OSWorld 2.0 offline subset, up from GPT-5.6 Sol's 65.7%, and it does the work in roughly 40 minutes per task versus Sol's 75. That's a 47% reduction in wall-clock time on the benchmark that most closely resembles what a small-business owner spends their afternoons doing.

The rest of the scorecard reads like the launch deck it's. Astra posts 98.6% on ARC-AGI-3 (Sol scored 7.8%; Claude Opus 5's baseline is 30%; NVIDIA's Agentic Variation Operators wrapped Opus 5 with persistent memory and tool orchestration to hit 100% in August). It scores 97.6% on FrontierMath Tier 4 v2. Greg Brockman calls it the start of the AGI era. Sam Altman, Jakub Pachocki, Aidan Clark, Mia Glaese, and CFO Sarah Friar rotated through the announcement and the follow-on briefings, including the disclosure that Astra tripped internal thresholds under the OpenAI Preparedness Framework and shipped with additional cyber guardrails under the Daybreak program.

Rollout to ChatGPT Plus, Pro, Business, and Enterprise begins in the coming days. API pricing lands at $10 and $50 per million tokens, which continues the trajectory we covered when OpenAI cut GPT-5.6 Sol API pricing 20%.

For a founder running a business alone or with a handful of employees, the practical question isn't the benchmark. It's what happens Monday morning.

Here's the honest read. Astra is a genuinely better executor. Point it at a CRM update, a form, a research task, a first draft of outreach, and it'll get further, faster, with fewer wrong turns. OpenAI's framing is that you can now delegate work "at a much higher level." That's true, and it's the most useful frontier-model release for owner-operators since the computer-use category began.

It's also incomplete. Astra doesn't decide which prospect is worth contacting today, which content angle is live this week, or which distribution channel your competitors are quietly winning. It waits for a prompt. The decision layer, the part that determines whether the executed work matters at all, still lives with the owner, or with something built to occupy that seat. That's the gap LemonLime is designed to fill: it studies the company, the industry, and the competition, decides what customer-growth work is relevant, and delivers finished output proactively at 9:00 AM local time.

A sharper executor isn't the same as a decision. It's the same pattern we flagged when OpenAI's own agent swarm quietly built a hidden message board: capability compounds faster than judgment about what to do with it.

Sources

• https://www.axios.com/2026/09/03/openai-astra-gpt-6-agi-brockman
• https://venturebeat.com/technology/welcome-to-the-agi-era-openai-launches-gpt-6-astra
• https://fortune.com/2026/09/03/openai-debuts-gpt-6-astra-computer-use-greg-brockman-says-start-of-agi/
• https://www.nbcnews.com/tech/tech-news/openai-debuts-gpt-6-astra-security-measures-rcna595940
• https://www.cnbc.com/2026/09/03/open-ai-astra-gpt-6-cyber.html
• https://www.bloomberg.com/news/articles/2026-09-03/openai-rolls-out-gpt-6-astra-model-with-added-cyber-guardrails