OpenRouter says Ox Alpha reaches 8 trillion daily tokens

• Ox Alpha reached an 8 trillion-token daily rate five days after appearing on OpenRouter, according to OpenRouter's five-day update.
• OpenCode reported 26 trillion Ox Alpha tokens across four days in OpenCode's post, while a chart of OpenRouter data put OpenRouter volume at 11.6 trillion in the first three full days.
• Free access, a roughly 1 million-token context window, and text, image, and video inputs put the model straight into coding-agent workflows, according to Pi's announcement.
• Prompt retention creates a second story: a terms comparison juxtaposes the model page's no-training statement with broader language about collecting content for training and improvement.

The OpenRouter model listing still identifies only an anonymous third-party provider, even as it routes a huge free preview. A terms review from The New Stack found the same data-policy tension in the linked documents. OpenCode's public data site is one more observable surface, while Cline said in its 48-hour update that Ox had already reached 8% of all inference through Cline.

The headline figures use different clocks. OpenRouter's first usage update put Ox Alpha near 6 trillion tokens for a day before its five-day follow-up reported 8 trillion daily.

A three-day OpenRouter chart reported 11.6 trillion tokens, 2.6 times the next-largest launch in its dataset. OpenCode then reported 26 trillion processed tokens over four days in its post.

These are separate platform counters and time windows, rather than a combined total. The numbers make this a public load test for long-running coding harnesses.

The official model description frames Ox Alpha around long-horizon software engineering, sustained agentic work, and mixed text and visual context. Pi's post describes a 1 million-token window with image and video support.

OpenRouter's launch command, ori [your favorite harness] --model stealth/ox-alpha, puts the model behind an existing agent CLI. The Ori documentation describes that wrapper as passing through familiar agent commands while centralizing access through OpenRouter.

Rohan Paul attributed the scale to long contexts, retries, and repeated tool calls inside coding-agent tasks in his OpenRouter-volume analysis.

The surge ran across several free routes. OpenRouter and Pi advertised free access, while the Nous Portal announcement advertised limited-time free use and one quadrillion tokens of daily capacity.

A follow-up access reply said any Portal tier, including Free, could use Ox Alpha. Capacity is an advertised ceiling rather than a usage count, but it explains how the launch could absorb demand that would be unusual for a paid frontier endpoint.

The public launch listed no disclosed owner, as an early report noted. OpenRouter's model page says the anonymous provider retains prompts and completions, while saying those records are not used for training.

The broader Stealth Model Terms contain language about collecting User Content for Stealth Model training and improvement, the conflict documented in a terms comparison.

A malformed-request experiment through OpenCode surfaced a Java class path associated with Zhipu's API route. A malformed-request probe interpreted that as serving-layer evidence for a Zhipu-hosted, unreleased GLM-family model; OpenRouter has not replaced the anonymous-provider designation.

On a real Cline repository bug, Cline said both Ox Alpha and Fable produced correct fixes, but Ox used roughly three times fewer thinking tokens. The comparison showed Ox stating a root cause once before editing, while Fable stated variations of it seven times.

That is one bug-level comparison, rather than a general efficiency benchmark, but it puts a concrete mechanism behind the model's agent-workflow appeal.

An early 10-task DeepSWE snapshot gave Ox Alpha 80%, a result recorded in davis7's initial test. A later result landed near 63%; davis7's clarification says that run was also a subset, albeit a larger and more useful one.

One separate Emacs-codebase experiment generated 45 commits, after which a Fable review found 14 significant bugs and roughly a dozen smaller issues in a code-review report.