Ox Alpha AI: Features, Pricing, Benchmarks & Model Origin

Ox Alpha AI is a recently released stealth AI model with unusually ambitious published specifications and very limited information about who created it. It became available through OpenRouter on August 20, 2026, with a 1,048,576-token context window, support for text, images, and video, tool calling, and up to 131,072 output tokens. It is also temporarily available at no token charge.

What is missing is just as important. The developer has not been disclosed. No official model card explains the architecture, parameter count, training data, or underlying checkpoint. Independent investigations strongly associate Ox Alpha with the GLM model family from Z.ai, also known as Zhipu AI, but they do not establish that Ox Alpha is GLM-5, GLM-5.3, or any other exact model.

In this blog, you will learn what Ox Alpha AI is, how its features work, how long its free access may last, what independent benchmarks reveal, why researchers connect it to GLM, and what privacy risks this anonymous model presents.

What Is Ox Alpha AI?

Ox Alpha is a hosted stealth AI model distributed under the OpenRouter model ID stealth/ox-alpha. In this context, stealth AI refers to a model released without publicly identifying its developer or exact technical lineage during the preview period.

OpenRouter describes Ox Alpha as a reasoning model for coding, sustained agentic work, complex reasoning, and production workloads. It is operated by an anonymous third-party provider; OpenRouter states that it only routes requests and is not the model’s developer, owner, or provider. OpenRouter’s current model listing is the principal first-party source for its published capabilities.

No weights, technical paper, architecture diagram, or training report have been released. Consequently, claims about its parameter count, mixture-of-experts design, training compute, or dataset should be treated as unconfirmed.

Ox Alpha at a Glance

Why Ox Alpha Is Getting Attention

Three characteristics explain most of the interest: temporary free access, a one-million-token context window, and early reports of competitive coding-agent performance.

OpenCode announced on August 20 that Ox Alpha would be free “for the next week,” with near-unlimited usage subject to available capacity. That announcement also claimed capacity for 100 trillion tokens per day. The capacity figure is a platform claim, not an independently audited measurement. OpenCode’s announcement should therefore be read as an availability statement rather than a performance result.

The anonymous launch also encouraged attempts to identify the model through tokenization, API errors, context behavior, and multimodal processing.

How Does Ox Alpha Work?

The model’s internal architecture has not been disclosed. There is currently insufficient public evidence to establish whether it is dense or mixture-of-experts, how many parameters it contains, or how it was trained.

What can be described is its external operating model.

Reasoning

OpenRouter categorizes Ox Alpha as a reasoning model. Independent endpoint testing has also observed reasoning-control behavior resembling recent GLM APIs, including reasoning-effort settings.

This shows that Ox Alpha exposes a deliberate reasoning interface. It does not reveal whether its reasoning process uses a particular training method or architecture. Any more specific explanation has not been publicly confirmed.

Large-Context Processing

The listed 1,048,576-token window allows an application to submit unusually large prompts, code repositories, documents, conversation histories, and tool results.

One reproducible black-box investigation retrieved three hidden “needle” strings from prompts measured at 934,221 tokens and found a practical request boundary near 1.005 million tokens. That supports the advertised serving capacity, but three successful retrieval tests do not prove that the model reasons reliably over every part of a million-token prompt. The investigation’s repository publishes its measurements and limitations.

Agentic Coding

In an AI coding agent, the model does not edit a repository unaided. An agent framework presents files and instructions, exposes tools, executes the model’s tool calls, runs tests, and returns results for another iteration.

Ox Alpha’s published support for function calling makes this loop possible. Its agentic performance therefore depends on both the model and the surrounding framework, including the available tools, timeout, prompt, repository environment, and retry policy.

Multimodal Input

OpenRouter says Ox Alpha accepts text, images, and video and returns text. This could support screenshot-based debugging, user-interface analysis, visual documentation, and video understanding.

Audio is not listed as an accepted modality. No authoritative source has published detailed vision or video benchmark results for Ox Alpha.

Ox Alpha's 1M-Token Context Window Explained

A context window is the amount of tokenized material available to a model during one interaction. Ox Alpha’s listed capacity is exactly 1,048,576 tokens, or (2^{20}).

That number is a capacity limit, not a guarantee of perfect recall. Long-context quality depends on where information appears, how much irrelevant material surrounds it, the prompt structure, and the type of reasoning required. The listing separately specifies 131,072 maximum completion tokens, but it does not clearly establish whether the maximum input and maximum output can be used simultaneously.

For practical use, teams should test retrieval and reasoning on their own document or repository structure instead of assuming that advertised context length equals uniform comprehension.

What Can You Use Ox Alpha For?

Coding

Ox Alpha can generate code, explain existing code, propose patches, and respond to test failures. The available evidence is strongest for coding-agent workflows rather than unsupervised production deployment.

Large Codebase Analysis

Its context window may allow applications to supply more files, dependency information, schemas, and previous tool output in one session. Large input capacity can reduce aggressive file selection, although irrelevant context may still lower answer quality.

AI Coding Agents

Tool calling allows an agent to read files, search a repository, edit code, and run tests. Independent DeepSWE testing shows that Ox Alpha can sustain long tool-driven episodes, while also revealing intermittent tool-call-format failures.

Technical Documents

The model can be used to summarize specifications, compare APIs, extract requirements, or review lengthy technical records. Factual outputs should still be checked against the original documents.

Multimodal Tasks

Its listed image and video support may be useful for interpreting screenshots, diagrams, interface recordings, or visual bug reports. No public evidence yet establishes its reliability across all visual formats.

Who Made Ox Alpha?

Why Its Developer Is Unknown

OpenRouter says the third-party provider chose to remain anonymous during the preview. OpenRouter explicitly denies being the developer, owner, or provider.

No authoritative announcement from Z.ai, Microsoft, Google, or another model developer had claimed Ox Alpha as of August 24, 2026. Therefore, the answer to “who made Ox Alpha?” remains: This has not been publicly confirmed.

Evidence Linking Ox Alpha to GLM

The strongest independent evidence concerns model-family fingerprints:

• Token-count patterns closely match GLM-5-generation models.

• API validation behavior and error formats point toward a Z.ai-compatible serving stack.

• Reasoning controls resemble recent GLM interfaces.

• Video token consumption reportedly matches GLM-5V-Turbo across controlled samples.

These signals support a GLM-family hypothesis. They do not prove ownership because a different company could host, fine-tune, distill, or wrap GLM-derived technology.

Is Ox Alpha Actually GLM-5?

There is currently insufficient public evidence to establish this.

One extensive investigation identified the original GLM-5 checkpoint as its best behavioral fit, while another fingerprinting project found the closest tested match was GLM-5.3. These conclusions are not necessarily contradictory: related checkpoints can share tokenizers, API conventions, and base-model characteristics.

The most defensible conclusion is that Ox Alpha appears closely related to the GLM-5 generation. The exact weights, checkpoint, fine-tuning history, and operator remain unconfirmed.

Is Ox Alpha Based on GLM-5.3?

Possibly, but this should not be presented as fact.

Tokenizer Evidence

The open-source modelprint investigation ran nine infrastructure probes against 12 candidate models. Ox Alpha matched Z.ai’s GLM-5.3 on six of nine probes, including four of four tokenizer probes. GLM-4.7-Flash matched five of nine, while the tested non-GLM models matched no more than two. The modelprint repository includes the probe code and cautions that a fingerprint match identifies infrastructure or family similarity, not exact model identity.

A separate investigation reported a 44-of-44 tokenizer differential match to the GLM-5 generation, compared with 42 of 44 for GLM-4.x. Its best checkpoint-level interpretation was the original GLM-5, not GLM-5.3.

Tokenizer evidence is strong evidence of lineage. It is not a unique signature for a checkpoint, because several related models can share a tokenizer.

API and Infrastructure Evidence

Independent probes found Chinese-language validation errors, GLM-like reasoning controls, and error-code patterns compatible with Z.ai infrastructure. These are more useful than asking the model to identify itself, since system instructions can force any deployment name.

However, API behavior may originate in a gateway or hosting layer rather than the underlying weights. It can suggest who operates the infrastructure without proving who trained the model.

Behavioral Evidence

An independent analysis tested four controlled videos and reported that Ox Alpha used exactly the same visual-token budgets as GLM-5V-Turbo for each sample. The tests varied duration, frame rate, and resolution. The tested GLM-4.6V, Qwen, and MiMo endpoints produced different token counts. The published evidence report describes the method and results.

Behavior and writing style are weaker evidence. Similar response formatting, coding strategies, or refusals can be produced through post-training, prompting, or gateway policies.

What the Evidence Does and Does Not Prove

The evidence supports this cautious statement:

Independent technical evidence points toward the GLM family and a GLM-5-generation model, potentially combined with GLM-5V-related multimodal technology.

It does not establish that Ox Alpha is stock GLM-5.3, a hypothetical GLM-5.3 Flash, GLM-5.5, or an official Z.ai deployment. Those stronger claims remain speculative.

The public specifications also fail to match a single current Z.ai product exactly. GLM-5.3’s official documentation describes text-only input, a 1M context window, 128K output, and mandatory reasoning. GLM-5V-Turbo supports text, images, video, and files but has a 200K context window. Ox Alpha combines the 1M context profile with multimodal input.

Is Ox Alpha Free?

Current Free Access

As of August 24, 2026, OpenRouter lists Ox Alpha at $0 per million input tokens and $0 per million output tokens. OpenCode’s Go documentation also lists ox-alpha-free as free for a limited time.

This is temporary preview access, not a confirmed permanent free plan.

Free Trial Duration

OpenCode announced on August 20 that the model would be free “for the next week.” A literal reading points to approximately August 27, 2026, but the announcement did not specify an exact cutoff time or timezone.

OpenRouter has not published a model-specific end date. Its Stealth EULA says stealth models are available for a limited time and may be removed with or without notice. Access through OpenRouter and OpenCode could therefore change independently.

Is the Free Access Permanent?

No permanent free plan has been announced. The current evidence supports only a limited-time preview.

Ox Alpha Pricing: What Could It Cost?

Current Confirmed Pricing

The only confirmed usage price on OpenRouter is zero for both prompt and completion tokens as of August 24, 2026.

What Is Known About Future Pricing?

No reliable future price can currently be established. The anonymous provider has not announced standard API pricing, a subscription, volume discounts, or enterprise terms.

Possible Pricing Scenarios

Current Z.ai prices offer context, not a forecast. Z.ai lists GLM-5.3 at $1.40 per million input tokens and $4.40 per million output tokens, while GLM-5V-Turbo is $1.20 input and $4.00 output. Z.ai’s pricing page confirms those rates. Because Ox Alpha’s identity and operating costs are unknown, transferring either price to Ox Alpha would be speculative.

Ox Alpha Benchmark Results

No official Ox Alpha benchmark report has been published by its anonymous developer. The available numbers are independent evaluations.

Early 10-Task Evaluation

Developer Ben Davis was reported as the tester behind the early run. It used a deterministic ten-task DeepSWE subset selected with seed 0. The tasks covered typed variable binding, JSON-schema references, FastAPI deprecation headers, Helm manifest streams, feature-schema persistence, KaTeX array spans, JavaScript resource declarations, restored query state, bounded-memory file spilling, and persistent analysis caching.

Ox Alpha reportedly solved eight tasks and missed scc-bounded-memory-spilling and vulture-persistent-analysis-cache. The published comparison showed Fable 5 at 65%, GLM-5.3 and Grok 4.6 at 62%, and GPT-5.6 Sol at 52% on these tasks. However, the reference models received four attempts per task, while Ox Alpha had one binary pass or fail. The reproduced task table makes this difference explicit.

One task changes Ox Alpha’s result by ten percentage points. This was an encouraging exploratory result, not “Ox Alpha’s overall coding score.”

DeepSWE

DeepSWE contains 113 original long-horizon software-engineering tasks from 91 active open-source repositories. It covers TypeScript, Go, Python, JavaScript, and Rust in isolated environments with program-based verifiers.

The complete independent run used openrouter/stealth/ox-alpha, Pier 0.3.1, mini-swe-agent, Docker, maximum default reasoning, and one attempt per task. It ran from August 21 to August 22.

Ox Alpha completed 66 of 113 DeepSWE tasks in this reported evaluation, equivalent to 58.4%. This is a task-resolution rate under that configuration, not general “accuracy.” The full evaluation repository publishes its configuration, logs, and result decomposition.

Of the remaining tasks, 22 were near-misses that passed at least 90% of new feature tests, while two implemented the feature but introduced regressions. Eleven of the 113 episodes ended after three consecutive responses containing no required tool call. Five reached the agent timeout.

LiveCodeBench

LiveCodeBench measures code generation against programming problems and executable tests. The independent Ox Alpha evaluation used the release_v6 set of 175 problems, greedy decoding at temperature 0, one answer per problem, and no agent, tools, or repository scaffold.

Ox Alpha solved 49 of 175 tasks, or 28.0% pass@1. Pass@1 means that the first and only generated answer passed all tests. Results were 22/43 on easy, 16/52 on medium, and 11/80 on hard problems. The evaluation repository documents a 20-second timeout per test and notes that all 175 attempts produced executable code.

This result measures raw Python solution generation under a custom harness. It does not measure the same capabilities as an agent that can inspect files, run commands, and revise a patch.

Why Ox Alpha Benchmark Results Differ

The three percentages answer different questions.

The 80% result came from only ten repository tasks. The 58.4% result came from all 113 DeepSWE tasks using a tool-driven agent. The 28.0% LiveCodeBench result came from single-shot code generation without tools or iteration.

Results can also change with prompt scaffolding, reasoning effort, tool definitions, timeouts, repository environment, sampling temperature, number of attempts, and verifier rules. The early run itself demonstrates variance: it failed the SCC spilling task, while the full run solved it. The JavaScript resource-declaration task moved in the opposite direction, from an early solve to a near-miss in the full run.

Averaging 80%, 58.4%, and 28% would be scientifically meaningless.

Ox Alpha Task Reflection

These outcomes are more informative than a headline percentage. Ox Alpha can make substantial multi-file repository changes and often gets close to a complete solution. The failures also show why automated tests and human review remain necessary.

Ox Alpha Strengths

Ox Alpha’s clearest strengths are its documented long context, multimodal input, tool support, and ability to sustain lengthy coding-agent sessions. Completing 66 full DeepSWE tasks is substantive independent evidence of repository-level capability.

The full run also found that 90 of 113 tasks passed at least 90% of their new feature tests. That indicates many failures were close, although a nearly correct patch is still a failure when it introduces regressions or misses required behavior.

Temporary zero-cost access makes the model unusually accessible for non-confidential evaluation.

Ox Alpha Limitations

Its largest limitation is uncertainty. The developer, legal operator, model architecture, parameters, training data, and exact checkpoint are undisclosed.

Tool reliability is also a measurable weakness. Eleven full-run DeepSWE tasks ended because the model repeatedly failed to issue a required tool call. The 28.0% LiveCodeBench result further suggests that strong agentic performance should not be converted into a claim of uniformly strong single-shot code generation.

The free preview may disappear quickly, and no production service-level commitment has been published.

Is Ox Alpha Safe for Business Use?

The answer depends partly on the access route.

The OpenRouter Stealth EULA, updated July 6, 2026, also prohibits submitting certain sensitive information and says stealth access may be withdrawn at any time.

Because the public OpenRouter statements conflict, the safest interpretation is that training rights and retention boundaries are not sufficiently clear for confidential use. Do not submit API keys, credentials, regulated data, customer records, unpublished designs, or proprietary source code without an appropriate contract and internal legal or security review.

For evaluation, use public repositories, synthetic data, isolated environments, least-privilege tools, protected branches, automated tests, and human review.

Ox Alpha vs Other AI Models

The closest useful comparison is with the documented GLM products implicated by the forensic evidence.

This comparison shows why investigators suspect a GLM relationship: Ox Alpha combines GLM-5.3-like context and reasoning characteristics with GLM-5V-like modalities. It does not prove that Ox Alpha is either product.

Benchmark rankings across unrelated models should wait for controlled tests using identical prompts, tools, harness versions, attempts, and timeouts.

What We Still Don't Know About Ox Alpha

The following remain unconfirmed:

• The developer, owner, and infrastructure operator

• Whether Z.ai or Zhipu AI is formally involved

• The exact base model or checkpoint

• Whether it is GLM-5, GLM-5.3, or an unreleased derivative

• Architecture, parameter count, and active parameter count

• Training data, training method, compute, and safety evaluation

• Knowledge cutoff

• Permanent pricing and availability

• Model-specific enterprise terms and retention duration on OpenRouter

• Formal rate limits and service-level guarantees

• Official benchmark results

• Whether weights or technical documentation will be released

Should You Try Ox Alpha?

Ox Alpha is worth testing if you need long-context coding, agentic repository work, or multimodal technical analysis and can use non-sensitive material.

It should be treated as a preview, not a settled production dependency. Run it against your own representative tasks, preserve logs, inspect every change, and compare it with existing models under identical conditions.

Organizations handling confidential or regulated information should wait for clearer provider identification, contractual terms, and data-handling documentation.

Final Verdict

Ox Alpha AI is a capable but unusually opaque preview model. Its large context window, multimodal input, tool support, and full DeepSWE result make it technically interesting. Its anonymous developer, temporary pricing, uneven benchmark results, tool-call failures, and unresolved OpenRouter data terms prevent a stronger production recommendation.

Independent evidence makes a GLM-5-family origin plausible, possibly with GLM-5V-related multimodal components. It does not prove that Ox Alpha is GLM-5.3 or that Z.ai is the operator. Until an authoritative disclosure appears, that distinction should remain explicit.