Benchmarks

The claim is simple to test: the same agent, through Induction, completes the same tasks for less at the same pass rate. Public agent benchmarks test both halves on tasks we did not design.

Setup

Every run compares two configurations on the same tasks:

Direct: the agent calls the model provider directly.
With Induction: the identical agent, with only its base URL changed to point at the proxy.

Same model, same prompts, same tools, and the benchmark’s own grading in both cases. Each benchmark runs at three reasoning efforts of gpt-5.6-sol (medium, xhigh, max), so the comparison is visible as the model works harder, and longer.

How to read the numbers

Results are reported as the change relative to the direct configuration:

Cost: the change in total spend at the provider’s list prices.
Accuracy: the change in the benchmark’s own pass metric.

Induction sometimes spends tokens to save tokens, making model calls of its own to decide what context to keep. Reported costs include that overhead, so the comparison reflects what you would actually pay.

Across both benchmarks and every reasoning effort, cost drops 37–67%. Pass rates hold at baseline, and more land above direct than below.

The benchmarks

tau3-banking: conversational customer-service agents with tools, a policy, and a simulated user. 97 tasks.
Toolathlon: multi-step tasks in real software environments through MCP tool servers. 107 tasks.