Baseten announced that its Model APIs now support deferred tool loading for open models such as DeepSeek, Kimi and GLM. The change keeps tool schemas out of the system prompt and loads them only when needed. In Claude Code tests with three registered plugins, coding benchmarks became 62.5% cheaper and 26.2% faster. Agentic benchmarks also saw savings of about 51% in cost and 15% in time.
Why it matters
Developers can run AI agents cheaper and faster without changing their models or code.