Ox Alpha Benchmarks: Community Scores Decoded

Quick answer: Ox Alpha has no official benchmark scores. The OpenRouter model page for stealth/ox-alpha lists no intelligence, coding, or agentic benchmarks, and the independent tracker Artificial Analysis does not list the model (checked August 22, 2026). The numbers floating around — roughly 80% on a 10-task DeepSWE subset and 87.5% on "Kingbench" — are community-reported, small-sample, and not independently verified. If you're evaluating Ox Alpha, this article explains what those scores actually mean and what signals to look at instead.

If you've been hunting for Ox Alpha benchmarks, you probably hit the same wall: a new model appeared on August 20, 2026, people are talking about it on X and in the OpenCode announcement, but the moment you go looking for hard numbers, the official pages are silent and the community posts disagree on methodology. With a stealth model — one whose provider is anonymous during preview — that silence is easy to misread as weakness, or to paper over with leaderboard screenshots. This guide is based on the OpenRouter model listing, public API data, and public announcements, updated August 22, 2026. We have not run formal benchmarks ourselves; the goal is a framework for reading the claims you'll see elsewhere, plus the official data that actually exists.

The Official Scoreboard Is Empty — On Purpose

Ox Alpha launched on OpenRouter on August 20, 2026. The official listing is detailed where it matters for integration: a 1,048,576-token (1M) context window, up to 131,072 tokens of output, text/image/video input with text output, a preview price of $0 per million tokens on both sides, reasoning that is mandatory and defaults to maximum effort, and support for tools, tool choice, and structured output. The positioning line is explicit: "a reasoning model designed for coding, sustained agentic work, and production workloads," suited to "long-horizon software engineering, complex reasoning, and workflows that combine text with visual context."

What the listing does not include is benchmark scores. There is no intelligence index, no coding leaderboard number, no agentic score in the Performance section. Artificial Analysis — the independent tracker most people check first — does not track this model (checked August 22, 2026). And the details people usually need to contextualize scores — provider identity, architecture, parameter count, and post-preview pricing — are not disclosed during this preview.

That is worth stating plainly: the absence of official scores is a fact about the listing, not evidence about capability. Stealth models often skip benchmarks, partly to avoid being pinned down before a full release. The right question is not "why no scores?" but "what signal can I trust instead?" — which is what the rest of this article answers.

What the Community Is Reporting

Two numbers are doing the rounds:

1. Roughly 80% on a DeepSWE 10-task subset. A community test reports Ox Alpha at about 80% on a 10-task subset of DeepSWE (community-reported, not independently verified). The same small-sample comparison puts Fable at 65% and GPT-5.6 Sol at 52%. Note the methodology: this is a subset of the benchmark — ten tasks, not the full suite — so the gap between 80% and 65% is literally a handful of tasks either way.

2. 87.5% on Kingbench. A community article reports Ox Alpha at 87.5% on Kingbench, placing it second behind GLM-5.3 at 91.25% (community-reported, not independently verified). Kingbench is not a standard benchmark: it has no published, auditable methodology, and no independent lab runs it. Treat the number as directional at best.

There is also speculation, based on fingerprint, tokenizer, and video-encoder comparisons, that Ox Alpha may be a hidden multimodal variant of a GLM-series model (unconfirmed, pure speculation). We don't endorse that guess, and it shouldn't influence your decision — but it explains why some threads treat the model as "expected good at coding" rather than "proven good at coding."

How to Read Community Benchmarks: A Practical Framework

Most write-ups just repeat the numbers. Here is what they usually leave out — four questions to ask before any community score earns your trust.

Question 1: How big is the sample? A 10-task result moves roughly 10 points per task. One lucky or unlucky run swings the entire headline. Rule of thumb: treat any score based on a sample of fewer than 50 tasks as a signal, not a conclusion. It tells you a model is worth testing; it does not tell you where it ranks.

Question 2: What does the benchmark actually measure? DeepSWE is a long-horizon software engineering benchmark — the model has to work through a real repository over many turns to resolve an issue. That resembles agentic production work. Kingbench's methodology is unclear (non-standard), so you can't even say what it measures. A single-turn QA score would tell you almost nothing about an agentic coding model, and vice versa. Match the benchmark's workload to yours before you weigh the number.

Question 3: Who ran it, and was it independently verified? Community tests are not peer review. If the person running the test has a stake in the result — a vendor, an affiliate, a hype account — discount it. The test of a test is reproducibility: can you see the exact task set, the harness, and the model configuration? If not, the number is a claim, not a measurement.

Question 4: Are the comparisons apples to apples? Was the same task subset used for every model? Same temperature, same tooling, same number of attempts? The DeepSWE 80% comparison looks internally consistent, but it is still a self-selected subset run by one party. If any model in the table was run under different conditions, the whole table is void.

The one-line summary: community scores are useful for exactly one thing — building a shortlist. They are not useful for ranking models, and they are especially weak for a stealth model nobody has been able to audit.

Better Signals: What Real Usage Says

Here is the part of the official data that most benchmark-hunting posts skip. OpenRouter's Apps/Activity view shows real production traffic, and in the roughly two days after launch, Ox Alpha consumed about 657 billion prompt tokens and 7.95 billion completion tokens. The five applications driving the most traffic are all agentic coding tools: Hermes Agent (120B tokens), Claude Code (108B), Oh-My-Pi (93.5B), DeepSeek Harness with its multimodal bridge (84.8B), and ZCode (51.5B).

Why does this beat an unverified 10-task screenshot? Because it is aggregated over thousands of users running real, long-horizon workloads continuously. Agentic tools are the hardest stress test there is: they hold context across many turns, call tools, recover from mistakes, and keep going. If teams are routing 657 billion prompt tokens through a model in two days, that is evidence of usefulness under realistic load — not proof of leaderboard standing, but a far stronger signal than a 10-task claim.

Operationally, the official monitoring (P50 over roughly three days) reports throughput of 24 tokens per second, latency of 5.81 seconds, uptime of 99.99%, and availability of 99.14%. Ecosystem adoption confirms the trend: Ox Alpha ("Ox Alpha Free") is listed on OpenCode Go with a limited free window for the next week, near-unlimited usage that won't count against Go plan usage, zero data retention, and the same 1M context. That is a public commitment from a tooling vendor, not a rumor.

If your work is long-horizon coding and agentic tasks — the exact profile the usage data shows — you can try Ox Alpha free on glm5.app right now.

How to Benchmark Ox Alpha Yourself

Given the state of public data, the most reliable benchmark for your use case is the one you run. Concretely:

Pick 3–5 tasks from your real work. A refactor in a repo you know, a multi-step debugging session, a long agentic run with tool calls, or a task that mixes a screenshot or document with instructions — the model accepts image and video input.
Run the same prompt through Ox Alpha and your current model. Same instructions, same tools, same temperature. Don't change the setup between runs.
Score completion and quality, not vibes. Did it finish the task? Was the output correct on the first try? How much hand-holding did it need?
Stress the long horizon. With a 1M-token context window and reasoning that defaults to maximum effort, Ox Alpha is built for sustained work. A single-turn question will not tell you what it can do; a 20-turn agentic run will.
Tune it fairly. The API supports reasoning effort (max/high/low), tools and tool choice, structured output, temperature, and top_p/top_k — so you can compare under identical, reproducible conditions.

A few caveats so you calibrate expectations: the provider is anonymous during this preview, the free price window on OpenRouter and the OpenCode free week may change, and nothing in this article is a substitute for the official pages. What you're testing is the model as it exists today — which is the honest way to evaluate a stealth release.

If you'd rather not wire up an API key to start, chat with Ox Alpha in your browser on glm5.app — the free web entry point — and run your first task in the next five minutes.

FAQ

Does Ox Alpha have official benchmarks? No. The OpenRouter listing includes specs, pricing, and performance monitoring, but no intelligence, coding, or agentic benchmark scores. Artificial Analysis does not list the model (checked August 22, 2026).

How did Ox Alpha score on DeepSWE and Kingbench? Per community tests: about 80% on a 10-task DeepSWE subset (with Fable at 65% and GPT-5.6 Sol at 52% in the same small-sample run) and 87.5% on Kingbench, behind GLM-5.3 at 91.25%. Both results are community-reported and not independently verified; Kingbench is a non-standard benchmark.

Can I trust the community scores? As signals, yes; as conclusions, no. They come from tiny samples with no independent verification. Run them through the four-question framework above — sample size, what's measured, who ran it, comparable methodology — before weighing any number.

Is Ox Alpha fast? OpenRouter's monitoring (P50, roughly three days) reports 24 tokens per second throughput and 5.81 seconds latency, with 99.99% uptime and 99.14% availability. Note that reasoning is mandatory and defaults to maximum effort, so expect thinking time on hard tasks.

How can I benchmark Ox Alpha myself? Run your own real tasks head-to-head with your current model — same prompts, same tools, same settings. The fastest way to start is the free web chat on glm5.app; the API through OpenRouter (free during the preview) lets you test programmatically with tools and structured output.