What Is Ox Alpha?

OpenCode currently offers it as a free preview. The platform claims infrastructure that can serve up to 100 trillion tokens per day. Other specs include a 1-million-token context window, multimodal support, reasoning, high-volume usage, zero-data-retention options, and generation around 50 to 60 tokens per second.

Ox Alpha showed up through OpenCode with an odd mix: free access, a huge context window, multimodal input, fast generation, and what looks like massive token capacity. I ran billions of tokens through it and hit it with cybersecurity benchmarks. During testing, I processed over 3 billion tokens in a few hours: 2.6 billion cached, 44.4 million input, 17 million output. The reported cost stayed at zero. The early results suggest it's more than a throwaway experiment. It's not the strongest frontier model out there. But at $0, it's surprisingly competitive. For developers running automated evaluations, coding agents, long-context experiments, or large-scale model tests, that changes the math.

OpenCode vs. OpenRouter

Ox Alpha also shows up on OpenRouter, but the experience differs. OpenRouter rate-limits free models. At testing time, its docs capped free models at roughly 1,000 requests per day. I hit that limit after about 1,000 requests and 10 million tokens. Enough for casual chat. Not enough for agentic workloads, automated evals, or long coding runs.

OpenCode gives far more generous access. The model may be free on both, but the infrastructure and policies around it are not the same.

How Good Is It?

I tested it on an internal cybersecurity benchmark with multi-step exploitation tasks. These aren't just code-generation problems. The model has to reason about the environment, identify attack paths, drop dead ends, write or modify code, and push through several stages to the goal.

On many easy and medium tasks, Ox Alpha did well. It repeatedly solved lower-tier challenges where some other models stalled. That makes it a plausible daily driver for reasoning and software engineering.

Cybersecurity benchmarks also hint at general reasoning: success requires understanding an unfamiliar system, forming a strategy, testing assumptions, writing code, interpreting failures, adjusting, and completing dependent steps. Ox Alpha held up reasonably across that workflow.

It Is Not a Frontier Breakthrough

Don't overstate it. Ox Alpha didn't solve everything. Several hard tasks stayed unsolved, and on higher-difficulty challenges, stronger frontier models won. A model that's excellent for free isn't the same as the best model available. Its real edge is the combination: decent intelligence, huge usage allowance, zero cost. That makes the economics more interesting than the benchmark ranking.

A Clue About Its Training Data

One task stood out. It involved exploiting a recent React Server Components vulnerability that could lead to remote code execution. Newer models handled it much better. Ox Alpha scored zero. Likely explanation: the vulnerability wasn't represented enough in its training data. If that's right, the training cutoff isn't extremely recent. But benchmark performance alone can't pin down a cutoff, so treat this as an observation, not a conclusion.

Reasoning Quality

I ran it with reasoning at the highest setting. Looking through successful trajectories, it wasn't flailing. It analyzed the target, weighed approaches, rejected weak ones, wrote supporting code, checked intermediate results, revised its plan, and worked through multi-stage tasks.

That pattern matters for software engineering. Complex coding work follows the same shape: understand the system, investigate, test, interpret errors, iterate. So the security results suggest Ox Alpha could carry over to other domains.

A Potentially Strong Coding Model

The initial testing points to a model that suits developers needing lots of inference. Think software engineering agents, repository analysis, large-context code review, automated testing, model evaluations, coding benchmarks, research workflows, long document processing, multimodal development tasks.

Token cost usually caps these workloads. Running hundreds of millions or billions of tokens through a commercial model racks up real bills. Free access changes that. Even if Ox Alpha is somewhat weaker than paid leaders, running far more experiments can beat squeezing the last bit of intelligence out of each request.

The 1-Million-Token Context Window

That reported context window could help with large repositories, multiple docs, logs, architecture notes, long conversations, big datasets, agent trajectories. A bigger window doesn't guarantee the model reasons well across all of it. But it gives more room when building agents or tools that need lots of information at once. Combined with the near-zero token cost, it makes Ox Alpha an attractive thing to play with.

Multimodal Support

Ox Alpha also appears to take images and files, not just text. For coding work, that means reading screenshots, examining diagrams, interpreting UI designs, processing documents, debugging visual interfaces, understanding technical images.

Ox Alpha is one of the more interesting free AI experiments around right now. It's not a major frontier breakthrough, and early cybersecurity tests show clear limits on difficult or recent tasks. But that's only half the story. Its combination of respectable reasoning, coding, multimodality, a huge context window, fast inference, and very generous free usage makes it unusually appealing.

Ox Alpha doesn't look like a model that makes every frontier system obsolete. The economics are the interesting part. Consider a model that delivers good reasoning, solid coding, multimodal input, a 1-million-token context window, decent speed, and huge usage limits. Now price it at zero. That combination enables experiments developers would never try when every million tokens carries a direct cost, and for researchers, agent builders, benchmark creators, and engineers running long autonomous workflows, the difference is enormous.

The preview shows how much AI experimentation changes when inference is effectively free. Developers who watch every token suddenly can run billions at no cost. That opens a different class of workflows.

The Infrastructure Around Ox Alpha

The infrastructure around Ox Alpha is the odd part. OpenCode has publicly referenced capacity around 100 trillion tokens per day. To grasp the scale: OpenCode previously reported more than 22 trillion tokens processed during a single spike tied to another model. Infrastructure at this scale is brutally expensive. Offering enormous inference capacity, large context windows, multimodality, reasoning, and zero-data-retention options for free naturally raises questions about who foots that bill.

Who Built Ox Alpha?

Who built Ox Alpha? The source material doesn't say. Some speculate a major American AI lab. Google is one candidate. And the original tester floated a link to a future Gemini 3.5 Pro-class model. But nothing in the material confirms that. It could be Google, another lab, an infrastructure provider running an experimental system, or something else entirely. Until the provider steps forward, any guess is just a guess. The big question: who built it? Until the provider reveals itself, Ox Alpha stays a mystery, whether it turns out to be Google or something else.

Is It Worth Trying?

Testing also turned up a notable absence of strong refusals during cybersecurity work, and the model completed security-related tasks that some general-purpose systems would decline. For authorized security research in controlled environments, fewer unnecessary refusals is a genuine plus. That trait shouldn't be mistaken for overall quality or safety, though. Only run security tests where you have explicit authorization.

Is it worth trying? Based on early results, yes, while free access lasts. It suits developers who need large volumes of inference rather than peak performance on every prompt, and it could work as a daily driver for programming, reasoning, agent workflows, large-context analysis, automated evaluations, and controlled security research. OpenRouter users should watch the free-model rate limits closely. The experience there may be far more restricted than through OpenCode.

The quality of those capabilities needs deeper evaluation. But their presence alone makes the model more versatile.