OpenAI: GPT-5.6 Terra

OpenAI's own positioning for this row, in GitHub's launch note, is The balanced default. A strong all-round choice for everyday interactive and agentic coding.source, accessed 2026-08-28 — the middle tier, not the frontier one. The catalog disagrees about what that middle buys. As observed on 5 September 2026, Terra beats the pricier release right before it wherever Artificial Analysis scores them both: the coding index reads 76.7openrouter-models, last checked 2026-09-14 here against 74.9openrouter-models, last checked 2026-09-14 for openai/gpt-5.5, agentic 43.7openrouter-models, last checked 2026-09-14 against 37.3openrouter-models, last checked 2026-09-14. The coding margin is the thin one, under two points; the agentic margin is several times wider. The price margin is not thin either, with Terra listing $2.00per million tokensopenrouter-models, last checked 2026-09-14 input against GPT-5.5's $5.00per million tokensopenrouter-models, last checked 2026-09-14.

Off the catalog's own scoreboard the gap looks wider than that thin coding margin suggests. Reporting from the June preview carries two figures the catalog does not. The first: Terra outpaced GPT-5.5's 83.4% on TerminalBenchsource, accessed 2026-08-28 — beating, on that benchmark, the same model it undercuts on price. The second: 91.84% on OpenAI's internal capture-the-flag cybersecurity testingsource, accessed 2026-08-28, enough to put the mid-tier model across the same "High" cyber threshold all three GPT-5.6 models crossed. That preview reached roughly 20 organisations on 26 June 2026 at the US government's request, thirteen days before this row's own 2026-07-09source, accessed 2026-08-28 general availability.