GPT-5.6

OpenAI made GPT-5.6 generally available on 9 July 2026, and the name turned out to be plural: Sol, the flagship; Terra, the balanced tier; Luna, the fast and cheap one. All three share roughly 1.05 million tokens of context, 128,000 tokens of output, and a knowledge cutoff at 16 February 2026. Only Sol unlocks the "max" mode (more compute) and "ultra" mode (parallel agents).

The pricing grid is the part worth reading closely. Per million tokens: Sol at $5/$30, Terra at $2.50/$15, Luna at $1/$6, each positioned opposite a Claude tier. Sol holds the price of the older flagship GPT-5.5 while being more capable, which forces the comparison onto capability per dollar. Two billing details complicate that: cache writes bill at 1.25×, reads keep their 90% discount, and traffic past 272k tokens carries a surcharge near $10/$45. An agentic cycle's bill tracks ingestion, at a read-to-write ratio around 153:1, not generation.

The benchmark picture splits by terrain. Sol leads on Terminal-Bench 2.1 and the Coding Agent Index (91.9% in ultra) at about a third less cost per task than Fable 5, while Claude stays roughly 15 points ahead on SWE-Bench Pro. OpenAI published an audit the day before calling 30% of that benchmark broken. METR separately flagged record reward hacking on Sol, swinging its time-horizon estimate from 11 hours to over 270.