GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 —
GROK-46H 865 —
GEM-37FH 865 —
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 —
GPT-6A 820 —
KIMI-K3X 810 —
CL-FAB5H 787 —
CL-OP5H 764 —
CL-OP46H 742 —
CL-OP47H 733 —
GEM-38FH 676 —
CL-OP47 583 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 —
GROK-46H 865 —
GEM-37FH 865 —
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 —
GPT-6A 820 —
KIMI-K3X 810 —
CL-FAB5H 787 —
CL-OP5H 764 —
CL-OP46H 742 —
CL-OP47H 733 —
GEM-38FH 676 —
CL-OP47 583 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

AA-Briefcase Runtime Data: Claude Opus 4.8 Takes 23 Minutes Per Task, GPT-5.5 Takes 11

AA-Briefcase has moved from scorekeeping into operations math. The benchmark already tested long-horizon knowledge work across 91 private tasks: financial models, board presentations, design mock-ups, and strategy deliverables built from thousands of messy source files. The new data layer asks the question buyers actually care about: how long does one of these agents take to finish the work?

The answer is uneven enough to change model selection.

The Runtime Table

ModelAA-Briefcase signalTime per task
Claude Fable 5Benchmark leader, unavailable~28.5 minutes estimated
Claude Opus 4.8Highest-scoring available model~23 minutes
GLM-5.21261 Elo, top open-weight model16.3 minutes
GPT-5.5 xhigh1159 Elo, top-five overall~11 minutes
MiniMax-M31113 Elo, next open-weight model~26 minutes

The key comparison is Opus 4.8 versus GPT-5.5 xhigh. Opus remains the best available model on AA-Briefcase quality, but GPT-5.5 xhigh sits on the Pareto frontier: top-five Elo, about half the runtime, and better throughput for teams running batches of work rather than one-off demonstrations.

That makes the benchmark less like a leaderboard and more like a capacity planning input. A model that takes 23 minutes per task consumes twice the wall-clock budget before cost is even considered. For a 91-task full run, that difference compounds into many hours of queue time.

Tool Calls Are Not the Bottleneck

The surprising number is not the 23-minute Opus runtime. It is the attribution. Tool calls and execution account for only about 12% of total elapsed time. The rest comes from output verbosity, number of turns, and model output speed.

That matters because it shifts the optimisation target. Faster sandboxes, better browser automation, and lower-latency file tools will help, but they will not fix most of the runtime. The larger lever is agent behaviour: fewer redundant turns, tighter intermediate reasoning, shorter deliverables where the rubric allows it, and models that do not spend their budget narrating the work back to themselves.

AA-Briefcase also exposes a familiar frontier-model trap. More time is not reliably more quality. GLM-5.2 scores 1261 and takes 16.3 minutes. GPT-5.5 xhigh scores 1159 and takes 11 minutes. MiniMax-M3 lands around 1113 Elo while taking longer than its result justifies. The runtime curve is not a clean capability curve.

What This Changes

The first wave of agentic benchmarks made the market argue about task success. The next wave is going to argue about task economics: quality, cost, latency, throughput, and failure recovery in the same frame.

For coding agents, this was already obvious. A model that solves a pull request in 40 minutes is operationally different from one that solves it in four, even if both eventually pass the test suite. AA-Briefcase brings the same logic to analyst work, product work, and strategy work.

The practical takeaway is blunt: the highest-scoring model is not automatically the best production model. Opus 4.8 is still the quality pick. GPT-5.5 xhigh is the throughput pick. GLM-5.2 is the open-weight efficiency pick. Once long-horizon agents are doing real work at volume, those distinctions matter more than another single-point Elo gain.