Methodology
What is compared
One coding agent (Codex in the ChatGPT desktop app, one model, one reasoning effort) does the same task twice at the same time. The two runs differ only in the browser tool the prompt tells the agent to use to check its work:
| Side | Tool | How Codex sees it |
|---|---|---|
| ChatGPT Computer Use | The in-app browser of Codex (the bundled Browser plugin, cua_repl) |
Top-level tool, as the plugin ships it |
| Mirasim Computer Use | Mirasim Computer Use as an MCP server named cu (computer_repl) |
Inside Codex's code-mode exec, with its description given up front |
Before each conversation is opened, harness/tool-exposure.ts writes the matching tool setup into the Codex configuration, so each new conversation picks up its own side's setup; after the run the actual setup is read back from the session record and a mismatch voids the pair.
A pair
harness/run-pair.ts (Team Tasks, the simulated project) and harness/run-real-pair.ts (real projects, the open-source repositories) run one pair:
- Each side gets its own working copy: Team Tasks with the task's seed patch applied, or a real project at the base commit with its git history removed. Each copy starts as a fresh git repository so the change can be inspected later.
- Each side gets its own mock API (with the task's backend scenario: latency, failures, sign-in, roles…) and its own dev server with hot reload.
- A script opens a new Codex conversation for each side in the ChatGPT desktop app (a
codex://link with the prompt, then Mirasim Computer Use's macOS driver presses Trust and Send) and sends the prompt. Odd-numbered runs send the ChatGPT Computer Use prompt first, even-numbered runs the Mirasim one. - Multi-round tasks: when a side finishes a round, its next round is typed into the same conversation. Rounds of the two sides interleave; one script sends everything, one message at a time.
- When both sides have finished, each side is judged on its own working copy (
docs/judging.md), and the result is written toruns/<side>/<task>/r<run>.json.
Sending is shared machinery and is not measured. If sending fails (the prompt never arrives, a stray dialog covers the window, the in-app browser does not start, the tool setup does not match), the pair is discarded and both sides run again (scripts/batch.sh; the discarded results are kept as r<run>.error-<time>.json). In v1, pairs were rerun 23 times this way.
What is measured
Everything comes from Codex's own session record (the rollout file in ~/.codex/sessions/), read by harness/rollout.ts:
| Metric | Definition |
|---|---|
| Time | Sum over rounds of task start → task complete, from the record's timestamps. Waiting between rounds is not counted |
| Model requests | Number of model calls |
| Input / output tokens | Total token usage reported by Codex (input includes cached input) |
| Browser calls | Calls to the side's browser tool; inside code-mode exec, each call written in the code counts, including calls through helper functions the model stored and reloaded |
| Pass | The judge's verdict on the final state |
The time therefore includes the model's thinking, all tool calls and the browser, but not the prompt sender.
Schedule of v1
scripts/formal-v1.sh: runs 1 to 3 go through the 51 Team Tasks tasks and then the 9 real-project tasks; runs 4 and 5 repeat only the 7 interaction tasks ★ and the 3 long features. The ChatGPT desktop app is restarted at the start of each block and every 10 pairs (long sessions made its in-app browser stop working). 200 pairs, run between 2026-11-06 18:26 and 2026-11-07 16:11 (UTC+8).
Environment of v1
| Component | Version |
|---|---|
| ChatGPT desktop app (macOS) | 26.1103.31037 for every run. The runner did not record it yet; it was read from the app's own logs, which print the release in use: 26.1103.31037 from 2026-11-06 16:45 to at least 2026-11-07 16:10 (UTC+8), covering all runs. The app updated itself to 26.1103.41825 later on 2026-11-07 and to 26.1103.60933 on 2026-11-09. The runners now write chatgpt_version into every result |
| Codex | 0.167.0-alpha.3.2 (from each session record) |
| Model | gpt-6.1-sol |
| Reasoning effort | medium in 394 runs, low in 6 (read from each session record; see below) |
| Mirasim Computer Use | 0.1.0-rc.5, as published on npm (@mirasim/computer-use@0.1.0-rc.5) |
| Judging browser | Chrome for Testing 153.0.8010.12 for Team Tasks; the installed Google Chrome for real projects |
| Node | 24.19.0 |
| Machine | One Apple silicon Mac running macOS |
The 6 low-effort runs
The conversations are started through the ChatGPT desktop app, which applies its own reasoning-effort setting. In 6 runs the session record shows low instead of medium: comments run 3, type-scale run 1 and tldraw-11253 run 3 on the ChatGPT Computer Use side; qa-members-error run 2, qa-board-count run 2 and hover-ghost run 2 on the Mirasim side. Both runs of a pair were never affected together. The report gives the totals with and without these pairs; no change moves by more than one percentage point.
Changes during and after the runs
Judges were audited and improved after the runs began; every change was applied to both sides by judging all finished runs again. The list and the 14 verdicts it changed: results/v1/judge-changes.md. Harness changes after the start (prompt sending, browser-call counting, judge retry, a pnpm setting for real-project snapshots) are listed at the end of the same file.