GPT-6不只Astra！Sol内测结果曝光，速度快6倍

TL;DR - OpenAI is reportedly testing GPT-6 Sol as a faster, lower-capability counterpart to Astra, while disclosing that internal research agents now perform roughly 3.1 agent-workdays per researcher workday. The developments suggest rapidly accelerating agent-assisted AI research, alongside growing concerns about monitoring increasingly autonomous systems.

• In one unofficial zero-shot SVG test, Sol produced roughly 28,000 tokens in 3 minutes versus Astra’s 25,000 tokens in 19 minutes; these are leaked anecdotal results, not formal benchmarks.
• OpenAI says its agents handle coding, environment setup, evaluations, debugging, experiment analysis, and training oversight, with median daily inference use exceeding $600 per researcher at API prices.
• OpenAI describes these systems as “automated research interns” that can complete bounded, multi-day tasks under human direction, but researchers still make high-level planning and deployment decisions.
• OpenAI’s chief scientist reportedly warns that chain-of-thought inspection is becoming insufficient for monitoring tool-using, collaborating agents and advocates restraint until stronger safety standards exist.