Thinking Machines' Inkling 2 Small scores 31 on Artificial Analysis at $1.18 per task, but uses about 4x the output tokens

A cheaper model that does business tasks well is worth a look, as long as you count tokens, not just the price per million.

What happened

Thinking Machines released Inkling 2 Small, a mixture of experts model with 594 billion total parameters and 17 billion active. It takes text and image input and has adjustable thinking effort. It is available now through the company's Tinker platform with closed weights. Thinking Machines says it intends to release the weights under Apache 2.0 after external safety testing, and Artificial Analysis lists that date as unconfirmed.

On the Artificial Analysis Intelligence Index it scores 31, up 18 points from the first Inkling and matching Gemini 3.8 Flash (high). AutomationBench-AA jumped from 5% to 46%. On AA-Briefcase, which tests long business workflows with spreadsheets, presentations and memos, it scored 1527 Elo versus 1439 for Claude Sonnet 5.5 (medium). Pricing is $0.30 in and $1.00 out per million tokens.

The catch: it averages about 150,000 output tokens per index task, roughly 4 times its predecessor and far more than the about 31,000 for Claude Sonnet 5.5 (medium). It still costs $1.18 per task, versus $7.77 for Gemini 3.8 Flash (high) and $10.02 for Sonnet 5.5 (medium). Its hallucination rate fell from 61% to 27%, mostly because it answers fewer questions. Thinking Machines also says it does not always pick the best tools for large codebases and does not support audio.

My take

Two things make this interesting for client work. It is built as a base for fine-tuning, and its strongest numbers are on business workflow benchmarks, not just coding.

But price per million tokens is not your cost. A model that writes 4 times more tokens can eat its discount on short tasks, and long outputs also mean slower runs. Before switching anything:

1. Run it on 20 to 50 real tasks from the workflow.
2. Measure cost per completed task and time to finish, not list price.
3. Treat "answers fewer questions" as a feature only if your workflow handles "I don't know" cleanly.
4. Wait for the weights if self hosting or data control is the reason you want it.