NexAU-AHE Leads Terminal-Bench 2.0 at 84.7% — GPT-5.5 in a Better Harness Beats OpenAI's Own Agent by 2.5 Points
NexAU-AHE, submitted by china-qijizhifeng on May 14, now leads Terminal-Bench 2.0 at 84.7% ± 2.1. The agent runs GPT-5.5 — the same model powering OpenAI’s own Codex CLI, now ranked fourth at 82.2%.
The 2.5-point gap between first and fourth place, on identical model weights, is a direct measurement of what scaffold engineering delivers.
The Leaderboard
| Rank | Agent | Model | Org | Score |
|---|---|---|---|---|
| 1 | NexAU-AHE | GPT-5.5 | china-qijizhifeng | 84.7% ± 2.1 |
| 2 | LemonHarness | Multiple | Lenovo CTO Org | 84.5% ± 2.6 |
| 3 | Capy | GPT-5.5 | Capy | 83.1% ± 2.1 |
| 4 | Codex CLI | GPT-5.5 | OpenAI | 82.2% ± 2.2 |
| 5 | Polaris | Multiple | PolarisOps | 82.2% ± 2.8 |
| 6 | WOZCODE | Claude Opus 4.7 | WOZCODE | 80.2% ± 2.1 |
| 7 | TongAgents | Gemini 3.1 Pro | BIGAI | 80.2% ± 2.6 |
All five top submissions were recorded on May 14, 2026 — the same date. That’s not coincidence: it reflects the benchmark community’s coordinated push to publish results before a known evaluation window closed.
What Changed Since April
OpenAI Codex CLI reached 82.0% on April 23 — the first lab-proprietary agent to top Terminal-Bench 2.0 and the first time OpenAI’s own harness led the leaderboard. Within three weeks, three external scaffolds had exceeded it. None of them changed the model.
NexAU-AHE’s 2.5-point lead over Codex CLI on GPT-5.5 specifically measures how much headroom a carefully engineered scaffold still finds in a frontier model. The gap is meaningful: it represents roughly the same magnitude as the difference between Claude Opus 4.7 (80.2%) and GPT-5.5 in a standard agent loop.
The Multi-Model Question
LemonHarness at #2 and Polaris at #5 both use multiple models, not a single frontier provider. Neither edges out the best single-model GPT-5.5 scaffolds — LemonHarness is within noise of NexAU-AHE (84.5% vs 84.7%), and Polaris sits at the same 82.2% as Codex CLI. Multi-model routing provides rough parity with strong single-model scaffolding here, not a decisive edge.
The Frontier Cluster
The confidence intervals on the top five entries all overlap at ±2-3%. Statistically, no one has broken away. But the direction is clear: the cluster is now setting up around 83-85%, with GPT-5.5 as the primary ceiling-finder and scaffold quality as the tiebreaker.
WOZCODE (Claude Opus 4.7, 80.2%) and TongAgents (Gemini 3.1 Pro, 80.2%) sit 2-4 points behind the GPT-5.5 cluster. That gap may compress when Gemini 3.5 Flash and Claude Opus 4.8 submissions arrive — neither is yet represented at rank 8 or above.
What Terminal-Bench 2.0 Measures
TB 2.0 runs agents in a real bash environment with no retrieval augmentation, no UI scaffolding, and no special context injection. A model gets a task, a shell, and a time budget. The result is one of the cleaner available measures of raw agent-level reasoning and shell tool use.
The score compression at the top — all five finalists within 2.5 points — suggests the benchmark is approaching saturation for GPT-5.5-class models under strong scaffolding. A harder revision would be needed to differentiate this tier further.