CyBT-CTF: GPT-5.6 family

GPT-5.6 is a family, not one tier. Read terra, luna, and sol as sibling lanes against the updated-harness GPT-5.5 baseline — never as a single collapsed GPT-5.6 score.

Each model's best CyBT-CTF run, sorted by solve rate. Harnesses differ across labs (shown per bar and in the tooltip), so read this as a capability landscape, not a controlled head-to-head. GPT-5.6 configs are outlined in white.

Bars are sorted by solve rate on the blinded CyBT-CTF benchmark. terra medium is the flagship (highlighted); GPT-5.5 was rerun on the same updated harness so the gain reflects the model, not the harness.

Solve rate versus total model cost (lower cost is to the right). Marker size follows Splunk-call volume — luna searches the most; terra hits the best accuracy for the fewest calls.

Every model's best CyBT-CTF run on axes you choose. Pick what to compare — accuracy, cost, or speed — and what the bubble size means. GPT-5.6 configs are outlined; harnesses differ across labs, so treat position as a landscape, not a controlled race.

The complete field — each model's best CyBT-CTF run, ranked by solve rate. GPT-5.6 rows are highlighted. Harnesses vary by lab and are shown per row; only sanitized aggregate scores, cost, and timing are included.

Does GPT-5.6 solve these because it investigates, or because it has seen the answers? Running each variant on the public BOTSv3 benchmark with no tools shows how much it can recall from memory alone.

BOTSv3 is public, so these rows are shown separately from the blinded CyBT-CTF leaderboard. They are useful for a public benchmark sanity check, not for the headline ranking.

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