Kite Launches Marathon Adaptive Inference for Long-Running AI Agents, Supports 5 Models

According to BlockBeats, Kite, an AI payment infrastructure platform, launched Marathon, an adaptive inference infrastructure for long-running AI agents on July 22. Users can select completion windows (now, soon, later, or anytime) for each request, with longer wait times offering lower prices—the deepest tier provides up to 65% savings compared to real-time pricing. The service features OpenAI-compatible APIs and one-command plugins for Claude Code and Codex. Marathon initially supports five leading open-weight models: Kimi K3, GLM 5.2, DeepSeek V4 Pro, Qwen3.6-35B-A3B, and Nemotron 3 Ultra, all under unified pricing.

Kite previously raised $33 million in funding led by PayPal Ventures and General Catalyst. The Marathon launch marks a key extension from agent identity and payment networks to underlying AI compute infrastructure, aiming to enable agents' independent execution and autonomous payment capabilities.