The most telling thing about GPT-5.6 isn’t the score jump — it’s which score they led with. The launch headlines the Coding Agent Index and Agents’ Last Exam, not MMLU. When a lab reorganizes its own marketing around long-horizon agentic tasks, that’s the tell for where the money already went. Sol reportedly beats the prior frontier on Agents’ Last Exam by double digits, and the new Ultra mode coordinates four agents in parallel by default. The unit of progress is now the loop, not the token.

That shift changes what I care about when a new model drops. A 3-point bump on a terminal benchmark is nice, but the operational questions are different: does the tool-call error rate drop enough to cut a retry tier out of my orchestration? Does “Max mode allocates more compute” mean my latency budget is now non-deterministic per request? Ultra running four agents in parallel is great for a demo and a real problem for anyone who has to reason about cost and rate limits across a fleet of them. Parallel agents multiply your failure surface, not just your throughput.

The three-tier split — Sol, Terra, Luna — is the part I’ll actually use. Most production agent work isn’t bottlenecked by the frontier; it’s bottlenecked by the cheap model in the inner loop that classifies, routes, and drafts. A capable Luna at a fraction of Sol’s price does more for a real system’s economics than another point on a coding leaderboard. The HN thread is already picking apart the benchmark methodology, which is the right instinct.

If the agent index is the new headline number, how long before someone ships an eval harness good enough to make these numbers reproducible outside the lab that reported them?