One 30-minute shift.
What each agent kept.

Audit of the disputed $0.91 figure

The three primary runs recorded 18,785,672 input tokens: 18,437,120 cached (98.1%) and 348,552 uncached, plus 19,700 output tokens including reasoning. At Standard GPT-6.1 Sol rates, those counters price to $0.697104 uncached input + $1.843712 cached input + $0.197 output = $2.737816 across three shifts, or $0.912605 per shift. The arithmetic is valid under those assumptions; it does not verify the real charge.

With caching disabled, pricing the same recorded token counts at Standard rates gives $12.59 per shift. That is a sensitivity scenario, not observed spend. Each scored shift ran for about 31 minutes; token billing does not price idle or robot-execution time as continuous model generation. Native Codex credit billing has no separate cache-write charge; zero native write tokens do not establish zero API write fees. All six benchmark attempts price to $5.11 under the original recorded-token Standard scenario; coding and development work are excluded. Codex remains outside the ranked cost graph until comparable billed usage is available.

Deviations, stated first

Two checks dropped before the freeze: the scripted wrong robot could not be made to fail them, so 33 of 35 count. Seven framework episodes ended in error and were re-run: six in the first wave, when Google rate-limited the project for ten minutes mid-shift, and one in the second. All seven were re-run together, a decision recorded before any re-run result. No OmniLink episode errored. One preregistered sensitivity check could not be computed: it needed every episode to stay above 0.9× real time, and on a shared laptop every wave dipped below at some moment, OmniLink's episodes included.

Read this as a system comparison

The original seven configurations used Gemini 3.5 Flash on September 27. Codex ran later with a different model and a later simulator build, with three simultaneous simulators rather than seven. It received the same full robot tools and instructions. The published shift and earlier results were known to the integration developer; each tested Codex session started fresh and could not access the test script, grader or prior results. This is a retrospective extension, not a fresh blind test or an isolated comparison of reasoning models. No statistical superiority claim is added.

Simulation only, on the same Windows laptop (RTX 3060 Laptop GPU; Newton/MuJoCo physics on CPU). The original reply checks include keyword grading. All three Codex runs dipped below 0.9× real time on the original five-second-window metric, so they do not support a real-time-clean sensitivity comparison; timing conditions remain a limitation. Token usage is recorded; Codex used the owner’s signed-in account and billed dollar cost is unavailable.

Recorded model rounds count native token-usage notifications; interrupted calls may lack final usage. Without keyword-graded reply checks, Codex averaged 86.1%; without any reply check, 84.8% (both post hoc).

The initial wave had one scored failure and two infrastructure errors from overlapping native interruptions. A narrowly scoped transport correction was documented before rerunning the entire three-repeat wave with a new source freeze. Both waves and source versions are retained in the archive; no prompt or grading change was made.

What this comparison establishes

Claude Code used the same 30-minute shift, 33 scored checks, full robot tools, instructions, machine and engine binary as Codex. Three fresh native sessions ran concurrently in separate simulators, isolated from the repository, grader and prior results. The original seven used Gemini 3.5 Flash on September 27; the native sessions ran on September 30 with different models and runtimes. These are retrospective system comparisons, not a fresh blind or model-controlled test.

Claude’s mean is 9.1 percentage points below OmniLink’s historical 89.9% and 4.0 points above Codex’s 76.8%. Its repeat scores overlap Codex’s; with only three repeats each, no statistical superiority is established. The adapter developer was a Claude Code session on the same model and knew the published results. Participant sessions had no access to that context.

Timing differs: Claude’s worst five-second windows stayed at 0.973–0.982× real time; Codex’s dipped to 0.147–0.260×. This may favour Claude because fixtures and motion depend on simulator timing. Simulation only, on machine 9722d23d12a3 (Windows 11, Ryzen 5800H, RTX 3060 Laptop GPU), with Newton/MuJoCo physics on CPU. This is not a safety certification.

Post hoc sensitivity: excluding keyword reply checks, Claude averaged 90.3%; excluding all reply checks, 89.4%. These do not replace the primary 80.8% score. The frozen adapter’s partial output counter undercounts tokens; the table uses the final native cumulative modelUsage totals instead.

A fresh live run

Offline verification is available now. Running the shift again needs OmniSim, the benchmark source and your own OmniKey and model key. Benchmark source and frozen protocols are included in the downloadable evidence. Native Codex and Claude Code runs additionally require their corresponding CLI and signed-in account. Live runs incur provider charges and will differ in detail: the model is sampled, and the provider's latency varies.