Straight answers about Ox Alpha

Ox Alpha is an anonymous 'stealth' reasoning model that appeared free on August 20, 2026. It has a 1,048,576-token (1M) context window, accepts text, images, and video, outputs text, and is positioned for coding, sustained agentic work, and production workloads.

Officially, nobody has said. Its public listing states only that it is operated by a third-party provider that chose to remain anonymous during the preview. Independent fingerprinting points strongly at China's Zhipu AI / Z.ai and its GLM-5.x family (likely GLM-5.3), but no lab has confirmed it, and analysts stress that shared infrastructure is not proof of model identity.

Yes, during the preview it is priced at $0 input / $0 output across several providers. This is a preview subsidy, not a commitment. The free window was announced as roughly one week from August 20 (about August 27, 2026), but no exact cutoff or post-preview price has been published, so treat 'free' as temporary.

Mixed and disputed. Its viral headline was 80% on a hand-picked 10-task DeepSWE subset, but the same tester corrected that to roughly 63% on the full 113-task set (mid-tier, near GPT-5.6 Sol mid and Opus 4.8). Independent reproducible tests are weaker still: 28% on LiveCodeBench v6 and #26 on an OpenCode coding leaderboard. It performs notably better inside a real agent harness with tools than in plain chat.

No. The operator is anonymous and retains all prompts and completions, and you cannot identify, audit, or contract with them. Do not send personal data, secrets, credentials, private source code, customer data, or regulated data. Use only sanitized test data, and route anything you would not want logged through a separate, attributable model.

The model's public listing states the provider retains prompts and completions but does not use them for training. However, this conflicts with the broader Stealth Program agreement (which reportedly permits collection for training) and with a competing 'zero data retention' claim on another host. The terms are ambiguous, so follow the strictest boundary.

The context window is exactly 1,048,576 tokens (1M), a provider-published figure confirmed across multiple platforms. The maximum output is 131,072 tokens (~128K), and that cap includes reasoning tokens. The 1M input is separate from and much larger than the 128K output cap.

The easiest way is right here: chat on the homepage, open the Playground for a full workspace with saved conversations, or call the API from your own code (the model id is ox-alpha, and the endpoint is OpenAI-compatible, so most SDKs work by swapping the base URL). Ox Alpha is also available on several third-party AI platforms under their own aliases. Because it is a no-SLA stealth preview, build in a fallback model for anything important.

Multiple independent tells align with Zhipu's GLM: tokenizer counts matched GLM-5.3 exactly (with only a fixed +75-token wrapper), video-token consumption matched GLM-5V-Turbo, it rejects audio like GLM-5V, it shares GLM's 1214 and 1301 error codes, its emoji rate matches GLM/Qwen, and a leaked Java stack trace named Zhipu's own API path. Z.ai also has a precedent of anonymous previews (it ran GLM-5 as 'Pony Alpha').

The most common complaint is getting stuck in loops or stalling on long tasks. Others include weak frontend/CSS and visual reasoning, verbose output, overthinking at high reasoning effort, middling interactive latency (~11.6s median agent turn), a ~4.45% tool-call error rate, and 503 errors or dropped streams under viral load. It also has no audio input and no schema-enforced JSON.

No. As of August 22, 2026 it had no entry on Artificial Analysis (no Intelligence Index number) or LMArena/Chatbot Arena, and its own public listing shows no capability benchmark. Every circulating score is community-run and unverified. Its strongest genuine ranking is adoption-based: No. 2 on OpenCode within three days of launch.

Yes. It supports function/tool calling (tools and tool_choice), SSE streaming, JSON output via response_format (without strict schema enforcement), and a reasoning mode with adjustable effort. It is OpenAI-compatible, so most SDKs work by swapping the base URL. Note the reasoning-effort levels are reported inconsistently: AI/ML API documents low/medium/high, while one fingerprint report lists max/high/low with no 'medium.'

Possibly soon. Prediction markets leaned toward a lab claiming credit around the end of the free window (a Manifold market priced ~65% odds by August 27, 2026, and ~87% by August 30). Unofficial leaks and fingerprinting do not count as an official claim. As of August 24, 2026 no company had confirmed it.