Ox Alpha: Fact-Checking the Free Stealth Model Everyone Is Adopting

On August 20, 2026, a model called Ox Alpha appeared on OpenRouter under the stealth/ox-alpha slug, listed as a “stealth model” operated by a third-party provider that has chosen to remain anonymous. Within three days it had a 248-point Hacker News thread with 194 comments, a Business Insider story asking who made it, and - most tellingly - a place as the default model in real developers’ coding agents. This site’s author runs one of those agents: within days of launch it became the daily driver at maximum reasoning effort, and it has handled every codebase pointed at it since.

This post reconstructs the timeline, separates what is verifiable from what was rumor, and maps the four best places to use it while it costs nothing.

The timeline: four chaotic days

Thursday, August 20: Ox Alpha goes live on OpenRouter. Friday, August 21: OpenCode announces Ox Alpha would be free for a week with “near unlimited usage,” citing provider capacity of 100 trillion tokens per day. Benchmark chatter spreads through developer feeds. Saturday, August 22: Business Insider covers the mystery; Stripe CEO Patrick Collison confirms he tried it (“it’s very impressive”); competing origin theories circulate. Sunday, August 23: the original Hacker News thread is still collecting comments while analysts publicly walk back their confidence about who built it.

The adoption numbers on OpenRouter’s public dashboard tell the rest of the story: Nous Research’s Hermes Agent routed roughly 1.36 trillion tokens through the model in the first three days, Claude Code nearly 700 billion, and smaller harnesses like pi hundreds of billions more. Whatever Ox Alpha is, developers did not hesitate.

The 80% DeepSWE rumor, corrected

Here is the piece of history that deserves a correction before it fossilizes. In the earliest chatter, word spread that Ox Alpha had scored around 80% on DeepSWE - the agentic coding benchmark from Agentica and Together AI, the same team whose DeepSWE-Preview reasoning agent was trained with reinforcement learning from Qwen3-32B. That number mattered because of the company it would keep: DeepSWE-Preview itself hit 59% on SWE-Bench-Verified with test-time scaling, then the top score among open-weight coding agents. An 80% reading would have leapfrogged the entire open field by more than twenty points in one stealth drop.

The verification never came: no leaderboard entry, no paper, no provider confirmation. What survived scrutiny was quieter but still unusual - independent arena results placing Ox Alpha near the top of multi-agent evaluations alongside frontier systems, and sustained heavy real-world traffic. The honest summary: Ox Alpha is one of the strongest models available for free right now, and the specific 80%-on-DeepSWE figure was a rumor that outran its evidence. Treat every stealth-model benchmark you see the same way.

The tokenizer detective story

With nobody claiming authorship, developers went full forensics on the model’s outputs. The dominant theory: Ox Alpha comes from the lab behind GLM. Wccftech reported that Z.ai - which previously tested GLM-5 anonymously under the name “Pony Alpha” - was a leading candidate, because developers identified similarities between Ox Alpha’s tokenizer behavior and responses and GLM’s known models. A competing analysis argued the tokenizer evidence instead points toward Microsoft’s MAI family. By Saturday morning, the analyst Andrew Curran summed up the drift: people seemed “less sure of anything.”

That uncertainty is the honest state of public knowledge. What can be said from published metadata is narrow: OpenRouter’s API classifies the tokenizer generically rather than naming a family, the model accepts text, images, and video as input and returns text, and reasoning cannot be fully disabled on OpenRouter - the lowest effort settings are low, high, and max. Everything past that is inference from behavior, and it should be labeled as such.

What is actually verifiable

Stripping away speculation, here is the confirmed specification, consistent across OpenRouter’s API, Venice’s model catalog, and OpenCode’s documentation.

The context window is 1,048,576 tokens, with up to 131,072 completion tokens - a million-token working memory aimed at long-horizon software engineering. The model is reasoning-first: reasoning is enabled by default with selectable effort levels (low/high/max on OpenRouter; structured-output support varies by gateway). Tool-calling and JSON modes are supported; OpenRouter notes response_format works without strict JSON-schema enforcement.

Price: zero. OpenRouter’s own FAQ states you are not charged for prompt or completion tokens; Venice lists all three pricing fields at $0.00.

Privacy terms are unusual for a free model: prompts and completions are retained by the operator but stated to be excluded from training, governed by OpenRouter’s Stealth Model Terms. Venice labels its copy “anonymized.”

The four best ways to use it right now

Free does not help if access is complicated. Four routes stand out.

Nous Portal is Nous Research’s inference API, carrying stealth/ox-alpha in its catalog alongside hundreds of other models; it is also home turf for Hermes Agent, the single largest consumer of the model. Command Code exposes the model under the same stealth/ox-alpha identifier for CLI-style coding-agent use.

OpenCode Zen, OpenCode’s curated gateway, lists “Ox Alpha Free” (x-preview-f-free) with input, output, and cached reads all priced Free, backed by that advertised 100-trillion-tokens-per-day capacity for the free period. And OpenRouter - the original listing - offers one provider hop, full metadata, and the easiest way to compare it against everything else in one dashboard.

All four speak OpenAI-compatible APIs, so switching between them is a base-URL and model-slug change.

Why this drop matters

Two things make Ox Alpha more than a curiosity. First, the stealth-drop format is becoming a real distribution channel: an anonymous lab bought market attention with pure capability plus a zero price, and got a Business Insider story, a front-page HN thread, and default-slot adoption in days. Second, it reset expectations for what “free” means - a million-token reasoning model with tool calling, served at a scale measured in trillions of daily tokens, costs nothing until its operators decide otherwise.

Enjoy it while it lasts, keep your benchmark skepticism calibrated, and if you figure out who built it, the entire timeline is still waiting for that answer.