The Oxbox Survey

New this week

• Ox Alpha was unveiled as GLM-5.3-Flash and disappeared from the free tier. On 2026-08-26 Zhipu claimed the model and released it as GLM-5.3-Flash, reported as a 320B-A18B mixture of experts with MIT-licensed weights on Hugging Face the same day. Both of those are trade-press reports, not a Zhipu primary release: the architecture and licensing figures are the outlets' and this survey has not verified them. The wccftech headline additionally makes claims about training hardware and serving volume that are no part of this item. What the survey itself measured is the disappearance. Within a day stealth/ox-alpha vanished from OpenRouter's catalog, x-preview-f-free from OpenCode Zen's roster, and z-ai/glm-5.3-free from ZenMux. The model survives only as the paid z-ai/glm-5.3-flash at $0.075/M in and $0.25/M out; its card advertises 1.31M context and the endpoint serves 1.05M (measured against the archived catalog, 2026-08-27). Its whole stealth arc ran seven days: listed 2026-08-20, revealed 2026-08-26, gone by 2026-08-27.
• The Ox Alpha fingerprint was right. A community blog post reported tokenizer analysis pointing at the GLM family; the reveal — on the secondary sourcing noted above — says GLM-5.3-Flash. Issue 0.2's matched experiment, which produced evidence against the attribution, ran against z-ai/glm-5.3-free, a different sibling than Flash, so its negative result never bore on the actual identity.
• Three new free models are on OpenRouter: minimax/minimax-m3:free (1M context), minimax/minimax-m2.7:free, and inclusionai/ling-3.0-flash-fin:free.
• The three nemotron nanos expired exactly on their published 2026-08-24 date: nvidia/nemotron-nano-9b-v2:free, nvidia/nemotron-nano-12b-v2-vl:free, and nvidia/nemotron-3-nano-30b-a3b:free. Their listings are gone with them; the survey's 2026-08-24 snapshot preserves the rows. This is the prediction issue 0.2 made, and the expiration_date field's first confirmation.
• Both inkling endpoints quadrupled their context to 1M, and three endpoints' max output was trimmed to exactly 90% of context (dots-3-note-preview 512K → 461K, nemotron-3-super 262K → 236K, glm-5.2 256K → 230K), which looks like one router-side change, not three vendor decisions.
• OpenRouter's free-tier rate limits are unchanged: 20 requests/min, 50/day, and 1,000/day after a one-time $10 purchase. These are platform-level limits covering every :free model. Other venues' limits were not re-checked this run.
• The OpenRouter stealth slot is empty for the first time since the survey started; big-pickle on OpenCode Zen is the last stealth model standing.

Top models to try

What I'm pointing oxbox at this week. The only USE-grade evidence in the survey now belongs to a model that costs money.

z-ai/glm-5.3-flash · OpenRouter · cheap paid: $0.075/M in, $0.25/M out

The revealed Ox Alpha: the model behind issue 0.2's run-backed findings, and the real credential leak it found in oxbox itself. Same weights per the reveal; same served behavior is an assumption until run. See the transfer caveat under "How far to trust this".

minimax/minimax-m3:free · OpenRouter · free

New this week. 1M context, response_format, and a 944K completion budget, by far the largest in the tier. Completion budget is the axis Ox Alpha taught this survey to check first. Takes image and video input.

nemotron-3-ultra-free, mimo-v2.5-free · OpenCode Zen · free

Not yet reviewed on anything.

No --max-tokens flag is needed any more: oxbox's default is now 100,000 (oxbox#9, merged 2026-08-27), sized for reasoning models after issue 0.2's runs showed the old 32,000 default truncating a real review mid-sentence. A run that still truncates now sets truncated=true in the status file.

This issue ships a machine-readable manifest of the table above, oxbox-manifest-2026-08-27.json. ox --manifest (oxbox#10, merged 2026-08-27) takes the first permitted entry, skips paid entries unless --allow-paid is passed, and moves to the next entry only with an explicit --failover. The new default and every flag named here require oxbox 0.1.0, released today (brew install curtisgalloway/tap/oxbox, or the deb attached to the release); runs record their own ox_version, so artifacts say which floor they were produced under.

The models

A listing of what the catalogs show, not a recommendation. Cloaked (or stealth) models are listed under a codename with no lab attached. Named models say who made them. This week the OpenRouter stealth slot is empty for the first time since the survey started; one stealth model remains, on OpenCode Zen. Keep in mind my standard advice on the tradeoffs of cloaked models.

Stealth

The OpenRouter slot is empty. Ox Alpha's exit and reveal are covered above. OpenRouter's stealth page still shows it as the current stealth model; the live catalog has no such row. The catalog is the better evidence, since the page is marketing and lags.

big-pickle (OpenCode Zen). Still in the roster, still cloaked, last probed 2026-08-23 when every call returned 503 from its upstream provider. Not re-probed this run.

Named — OpenRouter, 21 free

Plus google/gemma-4-{26b,31b}, nemotron-3-super-120b, nemotron-3.5-content-safety, nemotron-3-nano-omni, inkling-small (also 1,048,576 now), laguna-xs-2.1, two Lyria audio previews, and openrouter/free.

Named — ZenMux, 6 free

All six are free nowhere else. z-ai/glm-5.3-free and sapiens-ai/agnes-2.0-flash were delisted this week; the GLM delisting lands in the same window as the reveal, consistent with Zhipu closing its free preview paths together.

Free here still needs a funded account. Any balance clears the 402 reject_no_credit gate, and calls cost nothing after that.

Named — Requesty, 12 free

No changes since issue 0.2, the only venue with a clean diff. Five are free nowhere else, and this is the only venue publishing data terms as structured fields.

mistral/leanstral-1-5 has the cleanest data profile in the survey; its one review run produced six findings and zero real defects (issue 0.2). No card; 200 requests/day.

Named — OpenCode Zen, 63 models

The roster shrank by one: x-preview-f-free, which was Ox Alpha. Confirmed free by probe and still present: nemotron-3-ultra-free and mimo-v2.5-free; neither has been reviewed on anything. The venue publishes no prices; see "How far to trust this".

What we ran

Nothing. No oxbox runs and no probes since issue 0.2 (2026-08-24); this is a catalog-only mid-week run.

How far to trust this

• Use your own judgement. I don't vouch for the accuracy of any linked reporting about models. In this survey I will always label things I actually tried myself vs. things that the router or model vendors say; make your own decisions about who to believe.
• Issue 0.2's evidence does not automatically transfer to the paid endpoint. Issue 0.2's confirmed findings came from the stealth/ox-alpha route; z-ai/glm-5.3-flash is a different route, even though its endpoint serves the same 1.05M context (its card advertises 1.31M; budget against the endpoint figure). Same weights per the reveal; same served behavior is an assumption until run.
• OpenCode Zen publishes no prices at all. Its catalog returns model IDs and nothing about cost, so the snapshot records every row's free as null: unknown, not zero, and not inferable from a -free suffix. The only models known to be free there are the ones actually called.
• expiration_date earned partial trust this week. Three endpoints expired precisely on their published date, the field's first confirmation. Ox Alpha's 2098-12-31 ended after seven days. The field is evidence, not a promise, in both directions.
• The catalog and the runs are different evidence. Every table above is measured from snapshots/; nothing here speaks to review quality, and no model gained or lost a quality verdict this week because nothing was run.
• Scripted runs should read ox's status file, not shell plumbing. oxbox#11 (merged 2026-08-27) added --output, which writes the answer only on success, and --status-file, a JSON record of status, finish reason, token counts, and a truncated flag on every exit. The standing pipefail warning from issue 0.2 retires with it.

Generator review

Triggers checked: 10. Fired: 3.

T1 — a previously cloaked model was claimed. Fired, for the first time. The attribution history now has one resolved case, and the fingerprint was right: tokenizer-count analysis said GLM family in issue 0.1, the reveal says GLM-5.3-Flash. The skill's honesty rule says community fingerprinting "has a poor track record"; this survey's own tracked record is now 1 right, 0 wrong. Proposed edit, not applied:

old: "Community fingerprinting of a stealth model's origin has a poor track record. Report the guess, report that it is a guess, and report the base rate of past guesses being wrong."

new: same, plus: "Track record within this survey: 1 right, 0 wrong (Ox Alpha → GLM family, resolved 2026-08-26). The industry-wide base rate stays poor; one hit does not change the reporting rule."

T4 — two or more entries delisted. Fired mechanically on two venues (4 from OpenRouter, 2 from ZenMux). The proposed consequence, a standing churn section, is satisfied by "New this week", which leads the issue.

T8 — the oxbox workflow changed. Fired inside this issue's own window: oxbox merged three changes on 2026-08-27, all driven by this survey's findings — the 100,000 default token budget (#9), --output and --status-file with a truncated flag (#11), and manifest support with probe mode by default and opt-in --failover (#10) — all released the same day as v0.1.0, the version floor this issue's advice assumes. Acted on in this issue: both standing footnotes (raise --max-tokens by hand; pipefail) are retired above. Still for the skill: the scope section's description of the oxbox workflow is stale, and the data-gathering table should add the status file as an Observed-tier source.

Not fired: T2 (rate-limit structure re-checked, unchanged), T3 (no new venue), T5 (no review benchmark), T6 (nothing run, so nothing to contradict the catalog), T7 (the stealth slot went empty this week; the three-week counter starts now), T9 (a stealth model graduating to paid is model churn under T1, not platform economics), T10 (part 2 empty for the first week; that counter starts now too).

Process notes, awaiting the user's call on the skill edits. The issue format moved again at the user's direction on 2026-08-27: title "The Oxbox Survey"; "New this week" leads, then "Top models to try" (which may include cheap paid models), then the catalog as a pure listing with no verdict markers; catalog tables appear even on a no-change week; model names and news claims link to their sources inline, on first mention; summary prose uses K/M figures and spells out abbreviations; every issue ends with a human-readable Sources section. Two skill edits follow from this once the format settles: the report-format block needs rewriting to this layout, and the marker table needs a rule for what evidence a cheap-paid recommendation requires. Acted on from this issue's review: oxsurvey now archives each venue's full raw catalog under catalogs/ so pricing claims stay checkable; the v0 manifest shipped with this issue was approved and merged the same day (see T8); and a draft of the standing cloaked-models advice page exists with the user as primary author. Version stays 2.0.0 until the user approves.

Sources

What this issue drew on, and where to check it.

• The venue catalogs for OpenRouter, ZenMux, Requesty, and OpenCode Zen, captured by ./oxsurvey. Every table in this issue comes from those captures, and the OpenRouter catalog was queried again by hand for the ox-alpha and glm-5.3 pricing after the delisting showed up in the diff.
• OpenRouter's stealth listings page, which still showed Ox Alpha after the delisting.
• OpenRouter's rate-limit documentation.
• News coverage of the reveal: wccftech and explainx, found with the web searches OpenRouter "ox-alpha" stealth model revealed MiniMax and "Ox Alpha" Zhipu GLM-5.3-Flash revealed confirmed August 26 2026. Also read: codersera's stealth-model guide, written 2026-08-23 before the reveal; it is the reference for the community fingerprinting above, not for the reveal.

---

This edition is published from its source file. The evidence behind it, the catalog snapshots and the manifests, is in oxbox-survey, and nothing from there is edited on the way here.