Workflows written by a frontier model made an open model twice as fast.
We took workflows written by Claude Opus and had GLM-5.1, an open-weights model, execute them over 104 real one-day agent tasks. The workflows keep the model on task: run time fell by half in both runs, net score rose 4.5 percentage points on average, and zero-score failures dropped 17–42%.
| 104 tasks | Wall-clock | Net score % | Zero-score |
|---|---|---|---|
| Raw | 13.4h | 70.9% | 12 |
| + Opus-authored workflows | 6.2h2.2× faster | 75.4%+4.5 pp | 8.5 |
Tyler Willis · June–July 2026 experiments · Published July 2026
Setup
AgentIF-OneDay is 104 tasks a competent person could finish in a day: build a report, wrangle a spreadsheet, plan a trip, stand up a small website. Each task carries its own rubric. A single judge, Gemini 3.1 Pro, scores every criterion pass or fail, and the net is positive points minus penalties out of a possible 636.
All runs used the same agent harness (Pi) and the same weights (z-ai/glm-5.1 via OpenRouter). The only variable was the process the model followed.
This tests whether frontier intelligence can be transferred to lower-cost, open-weights models. In the raw condition, the model got the task and went. In the other, it executed a set of DeepWork workflows (explicit steps and quality gates for that kind of task) authored by Opus, not by GLM-5.1. Neither model ever saw the judge's criteria or its scores. Opus wrote a workflow guide for each kind of job as it did the work; GLM followed the guides.
Results
Completed in half the time and scored higher.
| Condition | Wall-clock | Est. cost | Net score % | Zero-score | Timeouts |
|---|---|---|---|---|---|
| Rawcontrol · n=1 | 13.4h | $56* | 70.9% | 12 | 1 |
| Opus-authored workflowsrep 1 | 5.9h2.3× faster | $11* | 76.7% | 7 | 0 |
| Opus-authored workflowsrep 2 | 6.5h2.1× faster | $20* | 74.1% | 10 | 0 |
The time result reproduced cleanly; the score and failure gains appeared in both runs, at different sizes.
Context: on the same benchmark and judge, Claude Code running Opus 4.6 scored 71.2% raw and 76.7% under DeepWork (three runs each, different harness). The open model's workflow runs land within a few points at 27–47× less per task ($0.11–$0.19 against $5.18; the cost readout has both ledgers).
DeepWork cut the GLM run's bill by 64–80% across the two runs: $11 and $20 in billable tokens against the raw run's $56. Structure did the opposite to the frontier model, raising Opus's cost per task 2.8× and roughly doubling its per-task time. The asymmetry is the mechanism: structure spends extra tokens gating a frontier model; it saves the tokens an open model burns wandering. The cost readout has the full ledger.
run ledger
$ ls benchmark_runs/ # net score · wall-clock, from judge summariesfull104_..._glm51_raw_control_20260611_002 451/636 48246.7sfull104_..._glm51_opus_v23_20260610_001 488/636 21210.3sfull104_..._glm51_opus_v23_rep2_20260611_001 471/636 23243.4s# same model, harness, and tasks; workflow runs add Opus-authored workflows + generic runner controls.
Signal
Improvements were not universal across tasks. On the two-run mean, 31 tasks improved, 29 declined, and 44 didn't move. The lift is a net effect with asymmetric magnitudes: the average improving task gained 43 points of its 100-point task score, nearly double the average decliner's loss of 22.
The gains cluster where you'd expect a process to help: tasks the raw model under-organizes, where it drifts, summarizes what it would do instead of doing it, or runs long and times out. The losses cluster where the scaffold gets in the model's way on work it already handles cleanly. Structure helps most exactly where the model helps itself least. That distribution suggests a routing strategy: run the comparison once, see which tasks the workflows help, and use them only on those. The cost readout estimates what that would be worth on the frontier model.
Under the workflows, the raw run's timeout and its drift episodes do not recur, consistent with where the seven missing hours went.
Caveats
We trust the speed result; the rest is preliminary signal.
The halved wall-clock reproduced cleanly in both replications. The score edge, the failure reduction, and the cost delta all point the same direction, but each rests on two runs against one control. This is an early readout: more raw baselines, more workflow replications, a judge-variance study, and ablations of the workflow condition are queued before we claim more.
Why it matters
There is ample room to cut the cost and time of recurring processes by moving execution to open-weights models. A chunk of what looks like model capability is really process — planning before acting, checking outputs — and process is portable. A frontier model can write it once; an open model can run it cheaply from then on.
There's a second mechanism we didn't use here at all: DeepWork can revise a workflow from the mistakes of its own runs. The idea is that consistency improves and token spend drops the more you use it. Measuring that loop with the same rigor is on the list above.
This is the same shape we keep finding: in a sibling experiment on this benchmark, Claude Code running under DeepWork cut its own zero-score failures by 36%, the sturdiest number this benchmark has given us. The durable value of borrowed process is dependability: an agent that finishes faster and fails less under the time and cost limits real work actually has.
Two follow-ups are running now. The same protocol is repeating on the newly released GLM 5.2 and on MiniMax M3, to see whether the effect follows the approach or this one model. And GLM is writing workflows for GLM to follow, testing whether the author has to be a frontier model at all. If that holds, the whole loop runs on open weights.