Inkling 2 Small: Tune Reasoning Effort, Pay Less Per Answer

Thinking Machines shipped Inkling 2 Small on November 5, and the part most write-ups skip is the dial: a single reasoning_effort value that decides how long the model thinks before it answers. On a model priced at $0.30 per million input tokens and $1.00 per million output tokens, thinking tokens are the whole bill, so that dial is a cost control as much as a quality control.

The launch coverage was about benchmarks (a 594B-total, 17B-active mixture-of-experts model, ARC-AGI-2 at 81.7% per ARC Prize, reported via the AINews recap of the release). This guide is about what you can do with it today without a GPU: call it through Tinker's OpenAI-compatible endpoint, measure how accuracy, tokens, latency and dollars move as you turn the effort dial, and then build a tiny router that starts cheap and escalates only when an answer fails a check.