OX Alpha AI: Stealth Coding Model Explained

OX Alpha AI is a stealth coding and reasoning model built for long-horizon software work, agentic tasks, and multimodal input. It is available through OpenRouter and promoted by OpenCode as a free preview.

Table of Contents

OX Alpha AI Specs and Pricing

Field	Verified detail
Model name	Ox Alpha
Model ID	stealth/ox-alpha
Release date	August 20, 2026
Access paths	OpenRouter, OpenCode
Modalities	Text, images, video
Context window	1,048,576 tokens
Max completion	131,072 tokens
Preview pricing	$0.00 input, output, cache-read
Free preview duration	1 week
Throughput	23 tokens/sec P50
Latency	4.39 seconds P50
Uptime	99.99% / 99.46%
Retention policy	Prompts and completions retained by provider; not used for training

What Is OX Alpha AI?

OX Alpha AI is a stealth model aimed at coding, agent runs, and multimodal reasoning.

The model has a 1,048,576-token context window, a 131,072-token completion cap, and free preview access. Its scale suits large codebases, long prompts, and workflows that use images or video alongside text.

How to Access It

Use stealth/ox-alpha in OpenRouter or OpenCode to send text, image, or video requests.

Open OpenRouter or OpenCode.
Select stealth/ox-alpha.
Send a text, image, or video prompt.
Compare the output with your current coding or agent workflow.

What Early Tests Suggest

Early coding signals look strong. In a user-run test, Ox Alpha scored above 80% on DeepSWE, ahead of about 65% for Claude Opus 5 and 52% for GPT 5.6 SOL. Since DeepSWE measures long-horizon software work, that result points to real promise on sustained engineering tasks.

Public benchmark discussion places Ox Alpha alongside Claude Opus 5 and GPT 5.6 SOL, with best pass@1 results in the low 70s on public leaderboard discussions. Users also reported that Ox Alpha found two real bugs in a Python project.

The most useful test is still your own workload. Run it on repo-wide debugging, agent loops, and multimodal prompts, then compare the quality and speed with the models you already use.

Pros, Limitations, and Best Use Cases

Best for coding: The model is aimed at sustained engineering work and long agent runs.
Huge context: The 1,048,576-token window can cover large repositories, logs, and multi-file prompts.
Multimodal support: It accepts text, images, and video.
Free preview: Input, output, and cache-read pricing are all $0.00 during the preview.
Privacy terms: Prompts and completions are retained by the provider and are not used for training.
Main limitation: The builder is still anonymous, so vendor identity and long-term availability remain unclear.

Ox Alpha makes sense for developers who want to test long prompts, complex code changes, or agent workflows without paying upfront. The anonymous release is the main reason for caution when vendor identity or long-term availability matters to your team.

Head-to-Head Comparison (Current Frontier Landscape, Aug 2026)

Aspect	Ox Alpha	Claude Fable 5 / Opus 5	GPT-5.6 (Sol etc.)	Grok 4.6	GLM-5.3 (public)
Context	1M	1M	~1–1.05M	500K	Similar (1M-class)
Multimodal	Text + image + video	Strong multimodal	Strong multimodal	Strong	Public version more text-focused; Ox appears multimodal-enhanced
Pricing	Free (preview)	Expensive ($5–10 / $25–50)	$2–5 / $12–30	$2 / $6	Paid
Coding/Agentic	Strong early signals (outperformed in small DeepSWE tests)	Usually tops or near-tops coding benchmarks	Strong, especially agentic	Competitive	Very close to Ox (fingerprint match)
Speed	Moderate (~23–29 t/s)	Varies (thinking adds latency)	Competitive	Often faster	Similar
Best for	Free long-context coding agents, multimodal agent workflows right now	Highest reliability on hardest coding/reasoning when cost is secondary	Balanced capability + ecosystem	Cost-efficient real-time / knowledge work	Chinese-lab ecosystem, GLM users

Ox Alpha coding benchmarks

Ox Alpha has no official coding or agentic benchmarks published as of August 22, 2026. OpenRouter lists none, and independent trackers like Artificial Analysis do not include it. All available numbers are community-run, small-sample evaluations that should be treated as preliminary signals rather than definitive rankings.

Reported Community Coding Results

1. DeepSWE 10-task subset (most cited)
DeepSWE is a strong, contamination-resistant long-horizon software-engineering benchmark (113 original tasks across 91 repositories and 5 languages). Community tester Ben Davis (and others) ran a deterministic 10-task subset:

Model	Pass rate (approx.)	Notes
Ox Alpha	~80% (8/10)	Near-misses on the 2 failures; one task (Marriott) solved in a single shot where others scored 0
Claude Fable 5	~65%	—
GLM-5.3	~62%	—
Grok 4.6	~62%	—
GPT-5.6-sol	~52%	—

This is a small subset (not the full 113-task suite). Full public DeepSWE leaderboards (mid-to-late 2026 snapshots) show frontier models in the ~70% range under full evaluation with the mini-swe-agent harness (e.g., Claude Opus 5 / GPT-5.6 Sol variants around 70–73% at max effort). The 10-task gap is therefore meaningful but carries high variance—literally a handful of tasks separate 80% from 65%.

2. Kingbench (community custom suite)
One evaluator’s private “Kingbench” (80-point scale covering coding, reasoning, multimodal, and specialized tasks):

Ox Alpha: 87.5% (70/80)
GLM-5.3: 91.25%
Claude Fable 5: ~82.5%
Qwen 3.8 Max: ~81.25%
Claude Opus 4.8: ~80%

Ox Alpha placed second overall and scored perfect or near-perfect on several coding/multimodal items (e.g., 3JS contact lens case, Panda SVG, Gemma fine-tuning task).

Other scattered signals

Early Terminal-Bench-style or agent-step counts mentioned in fingerprint analyses show Ox Alpha taking a high number of agent steps (similar to GLM-5.3 profiles).
Real-world usage on OpenRouter is heavily skewed toward coding agents (Claude Code, Hermes Agent, etc.), with hundreds of billions of tokens processed in the first 1–2 days—suggesting developers find it useful for sustained software-engineering workflows.

Context Against Established Coding Benchmarks (August 2026)

For reference, current public leaderboards (not including Ox Alpha) show:

SWE-bench Verified: Claude family (Opus 5 / Fable 5 / Mythos 5) dominates at 95–96%. Lower tiers sit in the high 80s.
Terminal-Bench 2.0 / 2.1: GPT-5.6 Sol leads (~90–92%), followed closely by Claude variants and Grok.
Full DeepSWE: Top models cluster around 70–73% at high/max effort.

Ox Alpha’s community subset scores place it competitively or ahead on the specific tasks tested, but without a full-suite run under a standardized harness it is impossible to rank it precisely against these numbers.

Analysis & Caveats

Positive signals

Strong performance on long-horizon, multi-file, agentic coding tasks in the tested subset.
One-shot success on a difficult task that defeated several frontier models.
High real-world adoption by production coding agents.
Matches the expected profile of a strong GLM-family model (tokenizer, style, step count, multimodal video handling).

Limitations

Sample size is tiny (10 tasks). Statistical noise is large.
Different harnesses, effort settings, and number of attempts (Ox Alpha often 1-shot vs. others 4 attempts) make direct comparison imperfect.
No public results on the full DeepSWE, SWE-bench Verified, Terminal-Bench, LiveCodeBench, or other standard suites.
Mandatory high-effort reasoning may inflate latency/cost once the free period ends, even if quality is high.
Some community voices have noted underperformance on simpler or non-agentic coding tasks relative to the hype, raising the usual “benchmaxxing vs. real-world” caution.

FAQ

Is OX Alpha AI free?

Yes. During the preview period, input, output, and cache reads cost $0.00.

How long is the free preview?

The free preview lasts one week.

Is OX Alpha AI used for training?

Prompts and completions are retained by the provider but are not used for training.

Who built OX Alpha AI?

Ox Alpha is a stealth model, and OpenRouter is the access layer.

How do I access OX Alpha?

Use stealth/ox-alpha in OpenRouter or OpenCode.

Bottom line: OX Alpha AI combines a huge context window with free preview access and strong early coding results. Try it through OpenRouter or OpenCode to evaluate long-context coding and agent workflows now.

FURTHER READING