← Back to feed

The current Terminal-Bench 2.0 leaderboard at tbench.ai carries three separate entries for GPT-5.5 — same model, different agent harnesses — spanning 2.5 percentage points: NexAU-AHE at 84.7% (±2.1), Capy at 83.1% (±2.1), and Codex CLI at 82.2% (±2.2). The base model is the same in all three cases. The scores differ because the scaffold differs.

The Harness Gap in Numbers

That 2.5-point range changes how leaderboard comparisons read. Claude Opus 4.7 running with the WOZCODE harness sits at 80.2% (±2.1). Gemini 3.1 Pro via TongAgents also lands at 80.2% (±2.6). When GPT-5.5 runs Codex CLI, it leads those models by roughly 2 points. With NexAU-AHE, the gap is 4.5 points. The model did not change; the quoted advantage did.

The confidence intervals make this more pointed. GPT-5.5 with Codex CLI at 82.2% ±2.2 and Claude Opus 4.7 with WOZCODE at 80.2% ±2.1 have overlapping error bars. A statement that GPT-5.5 definitively outperforms Claude Opus 4.7 on Terminal-Bench 2.0 requires specifying the scaffold — and the scaffold can be chosen after the fact.

Multi-Model Harnesses Offer Limited Lift

LemonHarness, which orchestrates multiple models rather than a single backbone, ranks second overall at 84.5% (±2.6) — barely beneath GPT-5.5’s peak. Polaris, another multi-model harness, sits at 82.2% (±2.8), tied with GPT-5.5 running Codex CLI. Combining models does not produce a step-change that sidesteps the single-model scaffold debate. The gains from multi-model orchestration are real but measured in fractions of a point, not tiers.

Why Scaffold Variance Reaches This Level

Terminal-Bench 2.0 tasks run in a bash environment against real codebases. What the scaffold controls matters: how the model sequences commands, maintains context across turns, handles errors and retries, and decides when to read files versus make changes. Each of those decisions is scaffolding behavior, not model capability. Two scaffolds running the same language model can produce meaningfully different task-completion rates because they implement different strategies for working in a shell.

The tbench.ai leaderboard displays Model and Agent as separate columns specifically to surface this. Rankings that collapse both into a single model-level comparison obscure a primary source of score variance.

What This Means for Model Selection

For teams using Terminal-Bench 2.0 to inform a model choice, the implication is that model and scaffold selection are entangled. A deployment running GPT-5.5 with a suboptimal harness may produce worse outcomes than a team running a nominally lower-ranked model with a more capable scaffold. The benchmark measures a (model, scaffold) pair. At the current frontier — where top model scores span roughly 4 points and harness variance spans 2.5 — scaffold optimization is where meaningful performance is being extracted, not model swaps alone.