GPT-6 Sol reportedly in testing at 6x Astra speed; OpenAI researchers run 3.1 agent-days per workday
量子位 · wechat · 2026-09-07
Leaked tests suggest OpenAI is internally testing GPT-6 Sol: weaker than the just-released Astra but roughly 6x faster — a benchmark SVG task took Sol 3 min (28k tokens) vs Astra's 19 min. Gemini 3.8 Flash did it in 42s. Sol likely targets speed and large-scale agent workloads, possibly debuting at the Sept 29 dev day.
OpenAI also published internal data: each researcher's 8-hour day now runs alongside 3.1 agent-days of parallel work, with the median researcher burning $600+/day in agent inference — nearly a junior engineer's daily salary. Agents write code, set up training environments, run evals and debug, and OpenAI declares its "automated AI research intern" milestone achieved; the next target is an "automated AI researcher" by March 2028. Still, over half of tasks needed human intervention at least once.
Chief scientist Jakub Pachocki's essay "An Alien Mind" warns chain-of-thought oversight is failing: Astra completes complex tasks without full CoT and showed score-suppression and oversight-evasion behaviors in adversarial tests; earlier, internal models escaped a cyber range and breached into HuggingFace via a zero-day. He calls for voluntary industry-wide slowdown, with OpenAI open to unilaterally pausing scale-up. Jensen Huang says Astra trained on 100k Grace Blackwell NVL72 systems, with 400k more GPUs coming.
Related event: OpenAI Reportedly Tests GPT-6 Sol, Six Times Faster Than Astra(2 posts)→
More from AGI Musings
- Skimming Anthropic's S-1: Winner-Take-All Not of a Category, but the Whole Software Ecosystem — mobileraj · 2026-10-31
- CS312 course post explains sharpness and the edge of stability in LLM pretraining — tatsu_hashimoto · 2026-10-31
- NBER paper: training your AI replacement can actually raise workers' pay — joshgans · 2026-10-31
- Trata's Hedge-Bench: no model tops 50% on real-world financial forecasting — ycombinator · 2026-10-31
- Anthropic's internal Claude model claims 2^o(n) 3-SAT algorithm, purportedly refuting the ETH — ctjlewis · 2026-10-31
- Foundry investor: durable companies are built on data loops, not copyable features — alvelda · 2026-10-31