CyBT-CTF: GPT-5.6 family

The best run for each model, with both harnesses shown for GPT-5.6. Every GPT-5.6 bar is outlined in white — solid for our stock Codex runs, dotted for Louie.ai-reported ones, which also carry the gold fill. Harnesses differ from lab to lab (each bar's label says which), so read this as a landscape of what's possible, not a controlled head-to-head.

Each model's best run, with cost shown per solved task. We place each GPT-5.6 config next to the frontier peers it lands closest to.

Bars are sorted by solve rate on the blinded benchmark. Terra (medium) and Luna (high) finish two solves apart, 32 and 34, and Luna spends about a third as much per solve. For reference we also show all three Louie.ai runs of the same models — Sol 45 of 59 (the top score on the board), Terra 34, Luna 31 — and every published GPT-5.5 lane: our updated-harness rerun plus the three original-harness runs (28, 20, and 14). All GPT-5.5 runs used high effort; no other effort level was ever run on CyBT-CTF.

Solve rate against total model cost (cheaper is to the right). Marker size tracks how many Splunk searches each ran — Luna searches the most, while Terra reaches near-top accuracy on the fewest.

The same best-per-model field as the frontier chart, on axes you choose. Pick what to compare — accuracy, cost, or speed — and what bubble size stands for. GPT-5.6 setups are outlined; the harness caveat above still applies.

Every published CyBT-CTF run — all harnesses and labs, not just each model's best — ranked by solve rate, with GPT-5.6 rows highlighted. We show only sanitized, aggregate scores, cost, and timing.

Does GPT-5.6 solve these by investigating, or because it has already seen the answers? CyBT-CTF is the clean primary benchmark (the blinded control is below). Running each variant on the public, contaminated BOTSv3 benchmark with no tools shows how much it can recall from memory alone.

Does the model take the task at all? We audited every GPT-5.6 with-tools run for outright refusals, and set the result against Fable — the one system on this board whose story is refusal — before and after its CVP unlock.

Sweep each sibling across its effort settings. Each peaks in a different place — Terra at medium, Luna at high, Sol at low. The chosen setting for each is highlighted.

Placed below the effort analysis on purpose. BOTSv3 is public and contaminated, so nothing here feeds the CyBT-CTF ranking.

Each line is one GPT-5.6 setup on the blinded CyBT-CTF benchmark — its solve rate, total cost, and how many Splunk searches it ran. Scoring is strict exact-match, so a right answer in the wrong format can still miss; read the scores as a floor, not a ceiling.