Manifests
Manifests
Each issue ships its picks as a JSON file that oxbox reads directly, so you can use the week's recommendation without transcribing it. One file per issue, oxbox-manifest-<date>.json, plus latest.json for the newest.
Current manifest
ox --manifest https://oxbox.ai/manifests/latest.json --files x.py --mode review "…"
latest.json is a byte-for-byte copy of
oxbox-manifest-2026-09-10.json, republished under a
stable name. Pin the dated file if you need a run to stay reproducible.
| Rank | Model | Venue | Cost | Why |
|---|---|---|---|---|
| 1 | z-ai/glm-5.3-flash | openrouter | paid | Editor's Rating Acceptable (2026-09-06): fast and mostly good; Z.AI's regulatory status is a concern. Six runs through oxbox on the three fixtures: ask-grounding 10 of 10 twice, the zero-defect control 1 real of 8 findings over two runs, the scanner fix passed all three gates as a baseline and produced a corrupt hunk as a candidate. Effort pinned at high, below the model's own default of max, because every survey run went out at high. OpenRouter's list price fell between the 2026-09-07 and 2026-09-10 catalogs, from $0.075/$0.25 to $0.07/$0.2333 per M, and max_price follows the new list. Pinned to the four routes those runs were served from; SiliconFlow billed 2x the catalog row on 2026-09-06, so max_price at list excludes it until it prices at list again, and with fallbacks off a refusal from the remaining routes fails the run instead of sending the payload somewhere unmeasured. |
| 2 | deepseek/deepseek-v4-flash | openrouter | paid | Editor's Rating Acceptable (2026-09-07): we fixed the routing problem; otherwise it's decent. The cheapest paid row in the survey. Nine runs through oxbox on the three fixtures: ask-grounding 10 of 10 twice; the scanner fix failed the apply gate twice; the zero-defect control 1 real of 4 findings over four candidate runs, the last two of them pinned to StreamLake and DigitalOcean by name, the first pinned runs the survey made. Effort high is the model's own default. Pinned to the three routes those runs were served from, with max_price at the catalog row: Novita billed 1.6x list on 2026-09-06, so the guard excludes it until it prices at list, and DigitalOcean's shared pool refused two of three pinned attempts on 2026-09-07 before answering, which with fallbacks off fails the run rather than sending the payload elsewhere. |
What is in one
A ranked list of models, each with the venue that serves it, the id to send, whether it is free, and the editor's one-line why. The list is derived from the Editor's Rating: models rated Good or Acceptable are in, Goods first and in the editor's order within a tier. Marginal and Poor stay out, so does any model with a standing disqualifier, and a model nobody has run cannot be rated and so cannot enter. Manifests dated before 2026-09-06 predate the Editor's Rating and were ranked by the old rule.
The format itself, and what oxbox does with an entry, are defined by oxbox, because it is the program that has to read the file. See Survey manifests in the oxbox README.
Stability
The dated files never change once published. latest.json is a byte-for-byte copy of the newest one and moves with every issue, so pin the dated file for a run you want to reproduce. Manifests are served with Access-Control-Allow-Origin: * and a five-minute cache, and carry a manifest_version; version 0 is the current one.