Ox Alpha is Z.ai's Answer to Fable 5 (Super SMART)
The short verdict
- Ox Alpha is worth a controlled test. In Ron’s Helm’s Deep build, it changed tactics when its first approach ran into trouble and finished with roughly 200,000 tokens. Comparable past channel runs often exceeded 1.2 million, although this was not a controlled model comparison. (source video hEGqHC2iDrY, 01:36; 03:19)
- The free deal is real today. OpenRouter currently lists
stealth/ox-alphaat zero cost for prompt and completion tokens. OpenCode Zen listsx-preview-f-freeas free too. Neither page promises a permanent free tier. - The Z.ai attribution is not confirmed. Ron believes the behavior could point to a new GLM checkpoint or a refined GLM-5V Turbo descendant. OpenRouter still identifies the developer only as an anonymous third-party provider. (source video hEGqHC2iDrY, 05:40; 06:44)
- Do not feed it private code yet. OpenRouter says the unknown provider retains prompts and completions, although it says they are not used for training. Ron therefore kept the model away from the channel’s private benchmark repository and GitHub OAuth. (source video hEGqHC2iDrY, 03:37)
Ox Alpha is interesting because it appears to spend fewer tokens repeating conclusions and more time trying a genuinely different repair. That could make it a valuable coding agent once the preview ends. One video and one external bug test are not enough to call it a Fable 5 replacement, however. Treat this as a strong reason to run a fair test, not as a finished leaderboard.
Watch the evidence
- 00:00 · First impressions and the Z.ai theory: Ron introduces the stealth model, the free window, and the first Helm’s Deep result.
- 01:02 · Why the reasoning felt different: Ox Alpha changes its calculation approach instead of repeating the same failed path.
- 02:20 · The Cline bug comparison: both models fix one real bug, but Ox Alpha produces roughly one-third as many output tokens.
- 03:38 · What the channel did not test: Ron explains why the private repository and the rest of the benchmark suite stayed disconnected.
- 04:25 · Community reports of better output: improving results lead to continuous-learning speculation.
- 05:40 · Checkpoint swap versus a new GLM: Ron favors the more ordinary explanation of a silent checkpoint or model swap.
- 06:20 · Try it before the preview changes: the practical recommendation and the unresolved provider question.
Ron in his own words
“When it comes across a problem, it doesn’t use the same approach to solve it.” — Ron, source video hEGqHC2iDrY, 01:03
“Only 200,000 tokens was used to finish the Helm’s Deep 3D project.” — Ron, source video hEGqHC2iDrY, 03:19
“You don’t really want to give it access to your sensitive stuff.” — Ron, source video hEGqHC2iDrY, 03:51
“I’m leaning more towards this. It could be the fact that it is a new GLM model.” — Ron, source video hEGqHC2iDrY, 05:49
What is confirmed, observed, and still speculation
| Claim | Evidence status | What you should conclude |
|---|---|---|
| Ox Alpha is free | Confirmed on August 25. OpenRouter shows free input and output; OpenCode Zen also labels its Ox Alpha route free. | Use the preview, but check the live listing before every large run. Free access, limits, and availability can change. |
| The model has a 1M-token context window and up to 128K output | Provider specification. OpenRouter publishes 1,048,576 context tokens and 131,072 completion tokens. | Useful capacity on paper, not proof that the model recalls every detail across a million-token repository. |
| It supports coding, tool calls, images, and video inputs | Provider specification. OpenRouter describes coding and sustained agentic work, plus text, image, and video input. | Test the exact harness, tools, and media format you plan to use. A supported modality is not a quality score. |
| It is a Z.ai or GLM model | Unconfirmed theory. Ron argues for it; OpenRouter does not name the maker. | Call it Ox Alpha or stealth/ox-alpha, not GLM-5.3, until the provider reveals it. |
| It is learning continuously from live traffic | Anecdote and speculation. Users reported different or better results on later days. | A checkpoint update, routing change, sampling variance, prompt difference, or backend model swap could produce the same pattern. |
| It beats Fable 5 | Not established. Cline reported one real bug where both succeeded and Ox Alpha used fewer output tokens. | Efficiency on one bug is a useful signal, not a general capability ranking. |
The ownership distinction matters. A product can be a credible alternative to Fable 5 without being built by Z.ai. Ron’s “Z.ai’s answer” framing is a hypothesis about capability and likely lineage, not evidence that Z.ai owns the route.
The efficiency signal is promising, not yet clean
Two observations make the efficiency story worth following.
First, Cline reported that Ox Alpha and Fable both fixed a real bug from the Cline repository. Cline said Fable repeatedly announced the same root cause before editing, while Ox Alpha stated it once and acted. The reported result was roughly three times fewer output tokens for the same successful repair. Ron connects that external result to what he saw in his own session. (source video hEGqHC2iDrY, 02:20)
Second, Ox Alpha used about 200,000 tokens on Ron’s Helm’s Deep project. He says earlier model runs on that benchmark commonly consumed 1.2 million or more. The agent also recognized a Three.js inconsistency and tried a different calculation instead of asking Ron to diagnose the problem. (source video hEGqHC2iDrY, 01:36; 03:14)
That is not a six-times efficiency benchmark. The video does not show matched provider conditions, identical harness versions, the same prompt and tools, repeated trials, equivalent visual quality, or a token-by-token audit. The free route was also under heavy demand, so wall-clock latency and model reasoning time were mixed together. Ron describes smooth tool use but slow thinking and loading. (source video hEGqHC2iDrY, 02:57)
The responsible reading is narrower: Ox Alpha showed a useful failure-recovery pattern and low token use in this run. Reproduce both on your own task before routing paid production work to it.
Why “improving every day” does not prove continuous learning
Community users saw stronger outputs after launch and asked whether Ox Alpha was learning from traffic in real time. The video explores continuous learning and fast adaptive layers, then lands on a simpler explanation: the operator may be silently updating a checkpoint or swapping the model behind the same route. (source video hEGqHC2iDrY, 04:25; 05:40)
There is no public training report, model card, version history, or controlled same-prompt record establishing online learning. The reference to Safe Superintelligence is a theory raised in the video, not evidence that Ox Alpha uses technology from SSI. Even repeated prompts are noisy evidence because provider load, hidden system prompts, sampling, tool responses, and route configuration can change.
If you want to test improvement, save the prompt, repository commit, harness version, tool permissions, reasoning setting, and acceptance tests. Rerun the same frozen task on different dates. A better screenshot is interesting; a repeatable increase in passing tests is evidence.
How to claim the free deal without exposing private work
The safest first run is a small public repository, a disposable local project, or a copied bug fixture with secrets and customer data removed.
- Choose a bounded task with a visible pass condition: one failing test, a small UI recreation, or a refactor with a fixed test suite.
- In OpenCode Zen, connect your account, run
/models, and chooseopencode/x-preview-f-free. The official Zen documentation currently lists its input, output, and cached reads as free. - Alternatively, use OpenRouter’s
stealth/ox-alphalisting, which currently shows zero-cost prompt and completion tokens. - Do not attach GitHub OAuth, a private repository, production credentials, internal documents, or personal data. OpenRouter says the anonymous provider retains prompts and completions.
- Record wall time, input tokens, output tokens, retries, changed files, test results, and the amount of human correction. Compare completed work, not the prettiest screenshot.
Free token pricing does not mean unlimited capacity, permanent availability, or zero privacy cost. It means the currently published prompt and completion rate is zero. Treat the preview as a chance to collect your own evidence.
Freshness and source boundary
The video was published August 25, 2026. This companion was checked the same day against the complete 1,326-word transcript and all 367 timestamp segments. OpenRouter and OpenCode’s live pages were also checked on August 25.
The article treats Ron’s Helm’s Deep run and interpretation as first-hand channel evidence. Cline’s bug comparison is a separate one-task report. Ox Alpha’s specifications, free status, privacy terms, and anonymous ownership come from current provider pages. No maker, GLM family, parameter count, training method, post-preview price, or exact preview cutoff has been confirmed. Recheck the live listing before sending a large job.