GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
← Back to feed

JJAgent Hits 87.1% on Terminal-Bench 2.0 — Multi-Model Routing Now Leads Every Single-Lab Agent

Terminal-Bench 2.0 is no longer a model competition. It’s a scaffold competition.

As of May 15, the leaderboard at tbench.ai shows JJAgent — a third-party agent from team JJ using multiple models — at 87.1%, ranking second overall. Only vix (90.2%, also third-party, running Claude Opus 4.7) sits above it. Below JJAgent: NexAU-AHE at 84.7%, LemonHarness from Lenovo’s CTO Research Lab at 84.5%, and Capy at 83.1%. All use multiple models or GPT-5.5.

OpenAI’s own Codex CLI, the lab’s direct entrant, ranks seventh at 82.0% — behind five other agents, four of which are from independent organisations.

The Leaderboard as of May 15–16, 2026

RankAgentModel(s)OrgAccuracy
1vixClaude Opus 4.7vix90.2%
2JJAgentMultipleJJ87.1%
3NexAU-AHEGPT-5.5china-qijizhifeng84.7%
4LemonHarnessMultipleLenovo CTO Org84.5%
5CapyGPT-5.5Capy83.1%
6PolarisMultiplePolarisOps82.2%
7Codex CLIGPT-5.5OpenAI82.0%

Four of the top six entries below vix use “Multiple” as their model designation, meaning they route adaptively across models per task. The JJAgent and LemonHarness results in particular — from a small team and a corporate research group, respectively — confirm that the technique is reproducible across organisations without access to frontier compute.

What Multi-Model Routing Gets You

The 5.1-point gap between JJAgent (87.1%) and Codex CLI (82.0%) using the same or equivalent underlying models is not explained by base model capability. It’s explained by scaffold architecture: how the agent chooses models, manages context, handles tool chains, and routes subtasks.

Terminal-Bench 2.0 measures real terminal workflow performance across categories including system administration, code compilation, server setup, data science, and cybersecurity. These tasks have heterogeneous requirements — some favour long-context reasoning, others favour fast code generation, others favour security-focused models. A scaffold that dynamically selects the right model for each subtask has a structural advantage over one locked to a single provider.

JJAgent’s 87.1% result — 5.1 points above OpenAI’s own Codex CLI — is the clearest quantification yet of what that advantage looks like at scale.

Why Labs Should Care

Model rankings at the frontier have treated Terminal-Bench as a proxy for model capability. The May 15 leaderboard challenges that assumption. A third-party JJ team without a foundation model beat OpenAI’s first-party agent on OpenAI’s own ground.

The commercial implication: scaffolding companies that nail multi-model routing — Augment Code, vix, and now JJAgent — may consistently outperform the labs’ own deployed products regardless of which underlying model is strongest. The base model still matters, but it’s no longer sufficient.

Terminal-Bench 2.0 now has 146 entries. The top 7 compress into a 8.2-point band (82.0%–90.2%). The next capability step requires either a substantially better base model or a substantially better routing strategy. Currently, routing is winning.