OpenAI GPT Astra
GPT-6.2 Astra now expected to score below 70% on AI exam, in a reversal
Traders who a day earlier expected a score of 70% or higher on HLE-Diamond now see about 68%; the reason is not yet clear.
Traders now expect OpenAI's new GPT-6.2 Astra model to debut below 70% on HLE-Diamond, a hard benchmark run by the Center for AI Safety, after treating a score of 70% or higher as likely only a day earlier. That outcome is now seen as unlikely, a collapse that came within a couple of hours on Tuesday.
Across the range of possible scores, traders' central expectation is 68%, down from over 70% when the story began, with a range from 64% to over 70%. A debut between 65% and 70% is the likeliest result. A score of at least 65% is still seen as likely, and a score of at least 60% is all but certain.
The reason is not yet clear. OpenAI launched GPT-6.2 Astra on Tuesday, which makes a qualifying model available, but the launch does not settle its HLE-Diamond result. No qualifying score from the Center for AI Safety has been found, and it could not be confirmed whether the model has appeared on the group's public chart or what it scored there. OpenAI reported results on other benchmarks, including 97% on GPQA Diamond, but that is a different test and does not count here.
The odds of a 70% debut had been steady and slightly rising for days. They sat at likely from Saturday through Monday, edging up on Monday, and held there through Tuesday morning. They began to slip around midday Tuesday and then fell steeply in the early afternoon.
HLE-Diamond is a version of the Humanity's Last Exam benchmark that the Center for AI Safety announced on Sept. 22. An earlier review put the predecessor model, GPT-6 Astra, at 59.9% under the center's setting of no tools and high reasoning. A 70% debut would be a large step up from that mark in a single model generation; traders now expect a smaller one.
Only the center's own figure counts: the accuracy shown for the model on its HLE-Diamond progress chart, read at noon Eastern on the day after the model first appears. Scores from OpenAI or other evaluators are excluded, as are tier-labeled and task-specific models.
The next development to watch is a chart entry or qualifying score for GPT-6.2 Astra from the Center for AI Safety. If the center publishes a qualifying score on the full 1,000-question test and the model is not charted within three days, the highest such score counts instead. A model must be added by June 30, 2027.
Sources
SiliconANGLE
OpenAI rolls out GPT-6.2 Astra for professional work and GPT-6.2 Sol for everyday work
Move history
- ▼ 10 ptsto 0.8%
- ▼ 62 ptsto 11%
Odds reflect traders’ expectations, not facts, and can swing on thin trading.