Astra's computer-use abilities and benchmark comparisons, in one long post
A long post explains that GPT-6 Astra can operate a screen like a person to complete tasks, and lists comparison figures including 72.6% on OSWorld and 41.4% on AutomationBench.
.@ChatGPT started as a chatbot. Four years later, it's basically an assistant that can use a computer like you do. Every model before GPT-6 Astra could talk to you about a task. This model can actually do it on your screen, the same way a person would. 𝗪𝗵𝗮𝘁 𝘁𝗵𝗮𝘁 𝗹𝗼𝗼𝗸𝘀 𝗹𝗶𝗸𝗲 𝗶𝗻 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗲: ➤ Fill out forms and update records without you touching a keyboard. ➤ Build and test spreadsheets or software on its own ➤ Browse websites, click around, and complete multi-step tasks end to end. ➤ Operate tools it was never specifically trained for, it just uses the screen like you would. 𝗪𝗵𝘆 𝗶𝘁'𝘀 𝗮 𝗯𝗶𝗴 𝗹𝗲𝗮𝗽 𝘁𝗵𝗮𝗻 𝗷𝘂𝘀𝘁 𝗮 𝗯𝗶𝗴𝗴𝗲𝗿 𝗻𝘂𝗺𝗯𝗲𝗿: ➤ On OSWorld (a benchmark for real computer tasks), Astra scored 72.6% in ~40 minutes per task. The previous model, GPT-5.6 Sol, scored 65.7% and took almost double the time. ➤ On AutomationBench, Astra hit 41.4% vs. 18.1% for GPT Sol which is more than double Against @AnthropicAI's current model, Claude Opus 5, @OpenAI's own numbers show Astra pulling well ahead on complex reasoning tasks (98.6% vs. 30% on ARC-AGI-3) 𝗪𝗵𝘆 𝘁𝗵𝗶𝘀 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗳𝗼𝗿 𝗿𝗲𝗴𝘂𝗹𝗮𝗿 𝗽𝗲𝗼𝗽𝗹𝗲 𝗮𝗻𝗱 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀𝗲𝘀: Normally, if you want AI in your software, you're waiting on that company to build it in. Astra skips that step and just uses the interface like a human employee would. So a business running old systems doesn't need to wait for an AI update because GPT-6 Astra can already work with it. 𝗕𝘂𝘁 𝗻𝗼𝘁 𝗲𝘃𝗲𝗿𝘆𝗼𝗻𝗲'𝘀 𝗰𝗲𝗹𝗲𝗯𝗿𝗮𝘁𝗶𝗻𝗴. Since the launch, a lot of the reactions hasn't been excitement... it's been worry. If the AI is doing the tasks instead of you, what happens to your own ability to do it? Majority are pointing to a few specific fears: ➤ We stop building patience because we never have to sit with a hard task anymore. ➤ Why learn to do something the model can do faster? ➤ Are we cultivating a generation that's great at prompting but weak at actually thinking through a problem? ➤ Is this a quiet shift from "using a tool" to "depending on a tool"? I think it's a fair tension, but whether that trade off is worth it probably depends on what you're using it for and how much you let it replace instead of assist. ✍️ https://x.com/OpenAI/status/2095595741528125780/video/1