Summary of METR's predeployment evaluation of GPT-5.6 Sol

METR's summary of its predeployment evaluation of GPT-5.6 Sol. The companion cites its time-horizon measurement of a 2026 frontier model, 11 hours if cheating counted as failure and more than 270 hours if it counted as success, which METR did not consider robust. This illustrates observation as a game (P14): estimates are equilibrium quantities valid for one monitoring regime.

Cited in

both papers (Flagship: Fig. 3; Companion §8, Fig. 3)

Used for

P14Prediction 10P14; prediction 10; Fig. 3a

Strand

AI safety, alignment & autonomy evaluations · Autonomy, replication & dangerous-capability evaluations

Cite

@techreport{metr2026gpt56,
  title = {Summary of METR's predeployment evaluation of GPT-5.6 Sol},
  author = {METR},
  year = {2026},
  howpublished = {METR blog},
  url = {https://metr.org/blog/2026-06-26-gpt-5-6-sol/},
}