On August 20, an AI model with no company name and no press release quietly appeared on OpenRouter under the identifier stealth/ox-alpha
. It has a one-million-token context window, accepts text, images, and video, and is completely free to use until roughly August 27. A developer ran it through a coding benchmark and recorded an 80% pass rate — ahead of Claude Fable 5 at 65% and GPT-5.6 Sol at 52%. Nobody officially knows who built it. That combination of performance, price, and mystery is why developer communities are paying close attention right now.
The Benchmark Numbers — With the Caveat Upfront #
Developer Ben Davis put Ox Alpha through ten tasks on DeepSWE, a software engineering benchmark designed to avoid contamination. Unlike SWE-bench, DeepSWE tasks are written from scratch across 91 repositories and five programming languages, then graded by behavior-based verifiers — not by tests that shipped with a specific merged fix. The point is to measure whether an agent can actually solve a new problem, not whether it has seen the answer during pretraining.
Ox Alpha scored around 80%. Claude Fable 5 scored 65%. GPT-5.6 Sol scored 52%.
Here is the caveat: Davis ran 10 tasks, not the full 113. A sample that small carries enough variance to make any result preliminary. Davis acknowledged this himself. This is not an audited leaderboard entry — it is one developer’s evaluation run. Treat it as a strong signal worth investigating, not a settled ranking.
The more reliable signal is what developers are actually doing with the model. Within the first days of availability, two major agentic coding harnesses had processed over nine billion tokens each through Ox Alpha: Claude Code (Anthropic’s terminal agent) and Hermes Agent (Nous Research’s open-source persistent agent). Production-grade tooling running billions of tokens is a harder thing to fake than a 10-task test.
How to Access It Before August 27 #
Ox Alpha is available through OpenRouter using any OpenAI-compatible SDK. The model ID is stealth/ox-alpha
. Pricing is zero — /bin/bash per token input and output for the duration of the preview. It is also accessible through OpenCode Zen (run /models
in the interface and select Ox Alpha) and Mercury Cloud.
from openai import OpenAI
client = OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="YOUR_OPENROUTER_API_KEY",
)
response = client.chat.completions.create(
model="stealth/ox-alpha",
messages=[{"role": "user", "content": "Refactor this function for readability: ..."}],
max_tokens=4096,
)
print(response.choices[0].message.content)
Context window is 1,048,576 tokens with a maximum output of 131,072 tokens. Throughput is approximately 24 tokens per second with a P50 latency around 5.8 seconds. That is not blazing fast, but for long-horizon coding agent tasks — cross-module refactors, architecture analysis, multi-file debugging sessions — latency matters less than reasoning quality and context capacity.
One important constraint: the provider retains prompts and completions. Since the developer’s identity is unknown, their privacy policy cannot be independently verified. Use Ox Alpha for open-source work, sanitized repositories, and evaluation spikes only. Do not feed it proprietary code or production secrets.
| Model | DeepSWE Score | Context Window | Cost |
|---|---|---|---|
| Ox Alpha | ~80% (10-task user test) | 1M tokens | Free (preview) |
| Claude Fable 5 | 65% | — | Paid |
| GPT-5.6 Sol | 52% | — | Paid |
Who Built It #
The strongest evidence points to Zhipu AI’s GLM-5.x series. Researcher Ben Davis documented four lines of technical fingerprinting: video encoder token consumption matched GLM-5V-Turbo across three independent encoder design choices, tokenizer alignment with GLM-5.3 held across 25 prompts with a consistent 75-token offset, audio input rejection followed the same pattern as GLM-5.x models, and output emoji usage rates matched Zhipu’s training data distribution.
Community testing suggests the capabilities may exceed GLM-5.3 — potentially an unreleased GLM-5.5 or a GLM flagship that Zhipu has not announced yet. Zhipu has not commented. DeepSeek, Xiaomi MiMo, Google, Qwen, xAI, OpenAI, and Anthropic have all been ruled out by the technical analysis. Check the comprehensive technical breakdown if you want to dig into the fingerprinting methodology.
This Is Now a Pattern #
Ox Alpha is the fifth stealth model drop on OpenRouter in 2026. The four before it were all eventually claimed by Chinese AI labs: Zhipu AI’s GLM-5, Xiaomi’s MiMo-V2-Pro (which went by “Hunter Alpha” first), Ant Group’s Ling-2.6-flash, and Meituan’s LongCat-2.0. The playbook is consistent: free preview to collect real-world feedback, technical fingerprinting by the community, then an official reveal tied to the commercial launch.
The fact that this keeps working as a strategy says something meaningful about the competitive landscape. Chinese frontier labs are now capable of shipping models that challenge the top commercial products from US labs — and doing it anonymously, without the usual marketing machinery. The gap has closed faster than most public benchmarks have reflected.
What to Do Before the Window Closes #
Run your own evaluation before August 27. The free period is specific and short. Check the DeepSWE leaderboard to understand how the benchmark is constructed, then run Ox Alpha against tasks representative of your actual workload — not borrowed benchmarks, but what your team does daily. If it holds up, you will have real data when the model either re-launches at a price or disappears.
Do not build production pipelines on it during the preview. Do treat it as one of the better free opportunities to stress-test a potential frontier coding model that this year has offered. If it vanishes on August 27, watch for a Zhipu announcement shortly after. The pattern is now predictable enough to plan around.