Delivered · Study13views

WindTunnel study - ab

webmcp vs playwrite · 50 tasks · 2 harnesses · 2 models

WebMCP delivers substantial efficiency gains—nearly half the cost and tokens, with a third faster execution.

Correctness did not separate the arms beyond the decision margin, and the treatment got there with 49% fewer tokens and 48% lower cost.

Abstract

WebMCP outperforms Playwright across all efficiency metrics in this 50-task benchmark spanning 2 harness/model cells. Cost came in at $0.06 (mean $0.10) versus $0.12 (mean $0.27), a -47.8% improvement. Token consumption was 282.5k (mean 436.3k) against 550k (mean 1.2M), representing -48.6% reduction. Duration improved -35.3%, with a median of 3m 55s (mean 4m 36s) compared to 6m 04s (mean 6m 28s). These gains held consistently across cells, though harness and model effects cannot be isolated separately due to the confounded design.

The result

playwritewebmcp

Best arm overall: webmcp (88.9 vs 79.1 of 100)

Best setup: claude-code / claude-haiku-4-5-20251001 · webmcp (98.5 of 100)

  1. 1claude-code / claude-haiku-4-5-20251001webmcp98.5best
  2. 2claude-code / claude-haiku-4-5-20251001playwrite88.2
  3. 3codex / gpt-5.4-miniwebmcp86.8
  4. 4codex / gpt-5.4-miniplaywrite76.0

Overall score per arm, 0–100 points (not a pass rate): 75% benchmark pass rate (the benchmark's own graders carry the outcome share — no judge grades or evals on this board) + 25% efficiency (cost · tokens · duration, vs the board's best arm). Absent components renormalize. Best at the top — the board's best setup is tagged.

Quality × efficiency clusters — normalized per task

Every graded run, standardized WITHIN its task so difficulty cancels out: → right = fewer tokens than the field on the same task, ↑ up = higher quality index (85% the benchmark's own pass/fail + 15% outcome) than the field. The field pools every harness and model, so a setup's left–right position largely reflects its own token habits. Dots are runs; each harness logo is a harness × model × arm centroid — hover it for the model and averages. Up-right wins.

Label
Task
Arm
Harness
Model
194 of 194 runs match
Show
playwritewebmcp
better · cheaperworse · pricier
← pricier than the fieldefficiency (σ, per task)cheaper than the field →

6 runs not plotted (missing a grade or token count — infra failures included).

Every metric, per harness × model

Benchmark pass rate

  1. claude-code / claude-haiku-4-5-2025100194% → 98%4.0 pp better
  2. codex / gpt-5.4-mini92% → 96%4.0 pp better

Task completion

  1. claude-code / claude-haiku-4-5-20251001100% → 100%even
  2. codex / gpt-5.4-mini100% → 100%even

Cost per run

  1. claude-code / claude-haiku-4-5-20251001$0.05 → $0.0350% better
  2. codex / gpt-5.4-mini$0.19 → $0.0953% better

Tokens per run

  1. claude-code / claude-haiku-4-5-20251001255.4k → 176.6k31% better
  2. codex / gpt-5.4-mini726.2k → 295.1k59% better

Duration per run

  1. claude-code / claude-haiku-4-5-202510013m 45s → 3m 30s7% better
  2. codex / gpt-5.4-mini7m 43s → 3m 59s48% better

playwrite

webdriver

webmcp

webmcp

The publisher has not described this setup — what each group's tools are, who built them, and how the tasks were chosen.

Distributions

Every completed run is one dot — the spread the averages hide. Click a dot to replay that run's journey.

Cost per run

playwritewebmcp
code/claude-haiku-4-5-20251001
codex/gpt-5.4-mini
0$7.90

Token composition — average per run

InputOutputCache readCache write
code/claude-haiku-4-5-20251001 · base
490.5k
code/claude-haiku-4-5-20251001 · treat
218.5k
codex/gpt-5.4-mini · base
2M
codex/gpt-5.4-mini · treat
663.4k

Run outcomes

How each run ended — whether the agent finished, not whether its answer passed. Whether answers passed is in the results above.

CompletedFailureTimeoutError / other
code/claude-haiku-4-5-20251001 · base
50/50 completed
code/claude-haiku-4-5-20251001 · treat
50/50 completed
codex/gpt-5.4-mini · base
50/50 completed
codex/gpt-5.4-mini · treat
50/50 completed

Statistics

MetricArmnMeanMedian (pooled)MinMaxStd dev
Costbase98/100$0.27$0.08$0.01$7.90$0.88
treat96/100$0.10$0.03$0.01$0.61$0.14
Durationbase100/1006m 28s6m 44s1m 03s22m 09s3m 43s
treat100/1004m 36s3m 49s1m 09s10m 40s2m 36s
Total tokensbase98/1001.2M360.3k34.4k34.2M3.9M
treat96/100436.3k178.1k26.1k2.6M585.8k
Output tokensbase98/1005.4k2.5k224114.6k13.3k
treat96/1002.3k1.1k33910.3k2.7k
Cache-read tokensbase98/1001.2M336.1k21.8k32.3M3.7M
treat96/100404k171.9k10.6k2.5M548.1k
Cache-write tokensbase98/1009k2.1k059.3k13.9k
treat96/1003.7k3.4k039.7k6.8k
Turnsbase100/10016.5829819.3
treat100/1009.983819

n = runs carrying the fact / completed runs in the arm; every statistic runs over present facts only — a missing fact is never counted as 0. Std dev is the sample form (n−1), withheld below n = 2. Every figure here POOLS all runs. The typical-run figures in the abstract, the head-to-head decision and the per-setup panels take each task's median first and then the median across tasks, so a task that ran more often never outweighs one that ran once; the two medians can differ. A promise study's headline (“cut average tokens”) compares per-arm means.

What each setup did

Derived from each run's recorded tool calls — not from a model's description of the run — and aggregated per setup, so a behavior seen across several runs is stated once with its rate. Open a finding to see the runs behind it, each linked to its journey at the step where it happened.

Harness
Model
Arm
Task
Pick a finding or a setup to list the runs behind it.
codex · openai · gpt-5.4-mini · playwrite3 findings
  • 5 of 50 runs · 5 completed
  • 5 of 50 runs · 5 completed
  • 5 of 50 runs · 5 completed

Task text withheld

Task text withheld — this benchmark is guarded and its tasks are not republished here.

Tasks

Runs

Every run is inspectable — open one to replay the agent's journey step by step, with the analyst's read underneath.

Every run, filterable… of 200Show runs
Task
Arm
Harness
Model
… of 200 runs match
Sort
HarnessModelTaskArmCompletedPassTokensCostDuration

Loading runs…

Methodology

What each metric means
Completed
The run finished and the harness returned a response. It does NOT mean the answer was correct — a completed run can score zero on quality.
Denominator: Terminal runs, excluding those killed by our own infrastructure.
Quality index
A composite used only in the quality x efficiency plot, built from whichever grader scored this study's runs — the judge's rubric, the pre-registered checks or the benchmark's own verdict — plus whether the run's own outcome was success. The map's caption names the grader and its weights.
Denominator: Graded runs — runs the study's grader scored.
Infra-excluded
A run killed by our own infrastructure. It is a missing measurement, never a loss for the arm, and is excluded from every rate denominator.
Denominator: Reported as a count beside every affected panel.
A/B arms
Every task runs twice per harness × model cell — once as playwrite, once as webmcp — on the same prompt, same model, cold start for both arms. Both arms are real configurations.
One run per task
Every task ran once per cell and arm, in one phrasing. Run-to-run variation is therefore not measured: a single task's difference can be one lucky or unlucky attempt. Every row of the reference table treats the TASKS as the unit: its intervals resample the tasks and its p-values come from a paired test over them, so they describe how much the result depends on which tasks were drawn — not how it would change if the same tasks were run again. With few tasks no difference can reach significance (five tasks cannot go below p = 1/16).
Sample size
2 cells × 50 tasks × 2 arms = 200 runs.