news-tldr.com

The important facts, the open questions, and the sources — without the churn.

← Latest briefing

Technology

Tests show GPT-6 Astra completing most block trials but struggling on puzzles

GPT-6 Astra succeeded in 19 of 20 bowl trials but stalled on puzzles, according to Hacker News.

Updated Sep 6, 2026 · 03:18 UTC1 reports

The short version

  • In testing evaluated by human graders on stages reached, GPT-6 Astra successfully placed a block into a bowl across 19 of 20 trials.[Hacker News]
  • On a puzzle insertion evaluation, GPT-6 Astra matched Fable 5.1 by completing only two out of 20 attempts.[Hacker News]
  • The model repeatedly reached the groove during the puzzle task but stalled at the same final step as Fable.[Hacker News]

Key facts

  • GPT-6 Astra successfully placed the block in 19 out of 20 trials during the bowl task.[Hacker News]
  • Human graders scored every trial according to the highest stage the run managed to reach.[Hacker News]
  • In puzzle insertion trials, GPT-6 Astra achieved completion two times out of 20, matching Fable 5.1.[Hacker News]
  • During unsuccessful puzzle runs, GPT-6 Astra reached the groove but stalled at the same final step as Fable.[Hacker News]

What remains uncertain

  • The underlying cause for why GPT-6 Astra consistently stalls at the final step of the puzzle task remains unspecified.[Hacker News]

Sources

Outlet counts describe coverage, not independent confirmation. Reports may share a wire service or original source.