Skip to main content
Benchmark overviewDetailed audit

Luna iOS: Stim vs local toolchain

Select a model, platform, scenario, and run to inspect its timing, commands, terminal output, and Settings-screen proof.

Settings-ready comparison​

Latest update 2026-09-11

JavaScript change

Stim reached Settings 73% sooner
Stim2m 24s
768k total tokens$0.033 total costRecorded 2026-09-1124 commands
Control9m 1s
2.59M total tokens$0.087 total costRecorded 2026-09-1132 commands
Benchmark machine

Mac mini​

Apple M4 / 16 GB

System
macOS 26.5.2 (25F84)
Toolchain
Xcode 26.6 (17F113) / Node 26.7.0
Simulator
iPhone 17 / iOS 26.5

Run audit​

Select a command bar to inspect its terminal output.

Settings ready2m 24sprimary outcome
Total tokens768kfull agent run, including cached input
Total cost$0.033full agent run / reported or API-equivalent estimate
Commands24708k cached input

gpt-5.6-luna / javascript / stim

Agent turn 3m 1s
What the agent did

Created an isolated worktree, worked on the JavaScript change, and ran Stim's iOS workflow. It reached Settings and captured valid agent-device proof. The record includes 1 failed command attempt before completion.

Agent / other 1m 50sShell active 1m 12sCommands summed 1m 12sPeak concurrency 3
0.0s / 3m 1s
1x
0.0s45s1m 31s2m 16s3m 1s
Agent
Shell
Events
Terminal25s / exit 0
$ stim worktree warm

  carry       copied .expo, android/.gradle, android/app/.cxx, +6 from worktree/bench-fixture-1.0.0
  carry       complete: 9 ignored entries copied, 0 kept, 0 failed
Commands in order (24)
Validated proof

Settings screen

Captured by agent-device after it found "Keep saved trail maps available offline". This screenshot completion is the timing endpoint used above.

Validated Settings screen for the javascript stim run
Run recording

Simulator playback

The run-scoped recording starts after the app session opens and includes onboarding and navigation.