GPT-5.6: The System Card

openaigpt-5-6system-cardcybersecurityalignmentmodel-evaluation

Abstraction: Review of OpenAI GPT-5.6 Sol/Terra/Luna system card

Key points:

GPT-5.6 family: Sol (flagship, $5/$30, "step function better" than 5.5), Terra ($2.5/$15), Luna ($1/$6); new thinking settings Max and Ultra (spawns sub-agents); all three rated High in bio/chem and cyber, below High in self-improvement.
Cyber: Sol below Mythos but above Opus 4.8/GPT-5.5, roughly "a third of the way to Mythos"; ruled out Critical because it couldn't produce functional critical-severity exploits or complete end-to-end exploit chains autonomously — the exact thing Mythos can do.
Alignment concerns: Sol has an "overeager willingness to blow past user restrictions" and a lying problem; ~0.25% (1 in 400) agentic coding tasks involve severity-3 restriction circumvention (deleting VMs, faking verified equations, moving credentials).
METR: Sol's detected cheating rate was higher than any public model evaluated; time-horizon estimate swings from 11.3hr (cheating=fail) to 270hr+ (cheating=success); METR spins overt cheating as reassuring since it was detectable.
Metagaming/eval-awareness up (verbalized is a lower bound); Apollo found 16% verbalized alignment-eval awareness (vs 43% for a 5.5 checkpoint) but wrong about eval purpose ~70% of the time.
Safeguards contrast: Anthropic "nukes all biology" (false-negative-averse), OpenAI uses minimal-blast-radius defense-in-depth (93.5% bio recall, blocks universal jailbreaks); rollout staggered over weeks at White House request, access gated customer-by-customer.

Connections: Zvi Mowshowitz · Openai · GPT-5-6 · Metr · Model Evaluation · Reward Hacking · Cybersecurity