Ox Alpha scores 80% on DeepSWE

Ox Alpha looks like an unusually compelling coding-model experiment, but its anonymity and limited evaluation data make the hype outrun the evidence.

• –The 80% result means eight of ten tasks passed, leaving substantial statistical uncertainty.
• –Its 1M-token context, multimodal inputs, tool support, and free preview are highly attractive for agentic coding workflows.
• –OpenRouter usage indicates immediate interest from coding agents, including Claude Code and Hermes Agent.
• –Provider identity, architecture, official benchmarks, and post-preview pricing remain undisclosed.
• –Privacy claims require scrutiny: OpenRouter’s general ZDR policy differs from the model listing’s reported provider-retention terms. OpenRouter ZDR documentation (https://openrouter.ai/docs/guides/features/zdr)