# November 5 matched performance comparison

The article's single six-agent tape uses this fresh, contemporaneous comparison, not a mixture of earlier campaigns. All six displayed agents ran on the same clean GCP VM, Codex OAuth account, native GPT-6 Luna at low reasoning, and 32-row ledger task. Native skills and extensions were enabled. Native prompts, tools, permissions and context budgets still differ.

[Task, fixtures and checker](fixtures/README.md) · [Full report](REPORT.md) · [Every main attempt](data/attempts.csv) · [Medians and ranges](data/summary.json) · [Frozen protocol](PROTOCOL-05.md) · [Schedule correction](PROTOCOL-05-ERRATUM.md) · [Checksums](SHA256SUMS) · [Download this snapshot](../matched-rerun-20261105.zip)

## What is shown

| Agent | Median time (s) | Median sampled peak tree RSS (MiB) |
| --- | ---: | ---: |
| octet 0.10.0 candidate | 9.465704470 | 38.289062500 |
| OpenCode 2.2.5 | 10.473369796 | 425.996093750 |
| Pi 1.5.0 | 10.954433434 | 144.765625000 |
| Codex 0.166.0 | 13.020968157 | 272.621093750 |
| fx 0.0.16 | 13.508250200 | 28.761718750 |
| OMP 18.16.0 | 23.773783933 | 388.410156250 |

Each metric has five measured attempts per client, after one excluded warmup. All five passed for every client in each pass; no retries replaced failures. Wall time is launch to exit without an RSS sampler. Memory is a separate pass using the observed process-tree sum at nominal 50 ms intervals, including runtimes and observable children. The chart converts KiB to MiB by dividing by 1024 and rounds to one decimal place. It does not use `wait4` maxrss or memory-pass timing to rank completion speed.

The downloadable records preserve the entire nine-arm cohort: the six displayed agents, Prime 0.10.2 with all nine bundled Python skills, and two historical Octet comparators. Prime and the historical comparators are not shown in the six-agent figure; they have not been removed from the dataset. The CSV contains all 108 main attempts: 90 measured runs and 18 warmups, including outliers and 13 recovered native tool errors. The report separately describes the earlier qualification failures; those qualification attempts are not main measured replicates and are not rows in this CSV.

## Exact Octet subject

- Source commit: `be5bcd39fa26ac024817dd82dfeb78fa99a894e6`.
- Executable SHA-256: `23a718110699df99ddbc568c92f2dd6aafda80b18b7a976e3b022d43805c0cdf`.
- Source-built 0.10.0 benchmark candidate, **not an attested final release artifact** or the later release-qualification tree.

## Interpretation

Octet has the lowest median wall time among the seven current clients in this task; fx has the lowest median sampled RSS. Five rounds do not establish a general or statistically resolved speed advantage. The five interleaved schedule rows are not fully position/predecessor balanced; the additive erratum preserves the original frozen protocol and execution order. Provider/network/cache variation remains uncontrolled.

Sampling can miss short peaks, double-count shared pages or miss descendants that detach before observation. The maximum recorded sampling gap was 63.641 ms. These numbers do not measure cold startup, streaming, provider RAM, general coding quality, long-context work or other models.

## Snapshot boundaries

The report, CSV, JSON, protocol, erratum, prompt, specification, checker and six ledger fixtures are byte-identical copies of the retained matched-rerun evidence. The medians and min/max ranges were independently reconstructed from the CSV when preparing the website. The original benchmark's separate replay reports agreement on 117 qualification/warmup/measured decisions; this website preparation did not rerun clients or claim to repeat that full native-session replay.

This small public snapshot excludes raw sessions, runtime installations, private profiles and credential stores. Paths mentioned inside the original report refer to the full retained benchmark directory, not files bundled in this download. Earlier benchmark campaigns remain available under [their original evidence directory](../performance-2026-11-05/README.md); they are not pooled with this rerun. The article's historical Terminal-Bench result and recordings retain their original versions and dates.