# Baseten — Deferred tool loading: Make your agent 62.5% cheaper without changing your model or harness

- Company: Baseten (baseten.co)
- Announced: 2026-11-19T17:49:14+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.baseten.co/blog/deferred-tool-loading-without-changing-model-harness
- Record: https://forck.live/items/38781-deferred-tool-loading-make-your-agent-62-5-cheaper-without-changing-your-model
- Subject: Baseten
- Models affected: DeepSeek, Kimi, GLM, GLM 5.3

Baseten added deferred tool loading to open models on its Model APIs, so agents stop paying for tool schemas they never use. The company says gateways serving open models had fallen behind because Codex and Claude Code turn deferred tool loading off when pointed at third-party open model providers, and it requires specific server-side gateway and inference work. Deferred loading keeps tool schemas out of the system prompt and gives the model a search tool; with three MCP servers registered, schemas added about 141k tokens to every request, which deferred loading cut to about 1.3k. Baseten says the technique is supported for DeepSeek, Kimi and GLM on its Model APIs in Chat Completions and Messages API formats. In Claude Code with Slack, Notion and Baseten MCPs on, coding benchmarks were on average 62.5% cheaper and 26.2% faster, while agentic benchmarks using the benchmarks' own tools were 51.1% cheaper and 14.9% faster. The measurements came from 55 tasks drawn from eight benchmarks, run with GLM 5.3 with and without deferred tool loading.

## Evidence

Verbatim from https://www.baseten.co/blog/deferred-tool-loading-without-changing-model-harness:

> The best way to cut costs and speed up your agents has always been hidden in plain sight.
> 
> Building harnesses have evolved from a discipline of addition to one of subtraction as we move into a world of lean, bitter-lesson-pilled harnesses. Edge cases and guardrails that used to contextualize the model for a task are now more often than not bloat.
> 
> So how do we let the model's judgment and intelligence shine through? One way to do this in practice is to retrieve tools just in time on a per-task basis, instead of preloading them, called deferred tool loading. This technique can also be understood as a form of progressive disclosure applied to tools to reduce wasted context.
> 
> Suppose you point your agent to any gateway so you can switch between frontier models and open-source models. In theory, this should work great. However, most miss deferred tool loading, which is typically supported by OpenAI and Anthropic APIs. Both Codex and Claude Code turn off deferred tool loading when pointing at a third-party open model provider (like Baseten).

---

Record: https://forck.live/items/38781-deferred-tool-loading-make-your-agent-62-5-cheaper-without-changing-your-model
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md
