New AI model from Thinking Machines matches a Google rival at a lower cost

Thinking Machines has launched Inkling 2 Small, an AI model that Artificial Analysis says matches Google's Gemini 3.8 Flash (high) on its main intelligence ranking. The cost gap is large. Running the test cost $1.18 per task for Inkling 2 Small and $7.77 for the Google model. Artificial Analysis also says it would be the top open model outside China once its weights are released.

What changed

Inkling 2 Small is a big step up from Inkling Small, the earlier version. On one test of automated tasks, its score rose from 5% to 46%. On a hard coding test it rose from 5% to 25%. It also beat Xiaomi's MiMo-V2.6-Pro on the same overall index.

• It costs $1.18 per test task, against $10.02 for Claude Sonnet 5.5 (medium).
• It costs almost 6 times more than Inkling Small, which was $0.20 per task.
• It uses about 150,000 output tokens per task, roughly 4 times its predecessor. A token is a small chunk of text.

Why business users may care

One test, AA-Briefcase, covers long business jobs that need spreadsheets, presentations and memos. Models get a ranking score where higher is better. Inkling 2 Small scored 1527, ahead of Claude Sonnet 5.5 (medium) at 1439 and Gemini 3.8 Flash (high) at 1203. It passed 50% of the quality checks, up from 20% for Inkling Small.

Its price per token is low too: $0.30 per million input tokens and $1.00 per million output tokens. That helps explain the low cost per task, even though it writes far more text per task than the others.

Where it falls short

It is not best at everything. On a terminal-use coding test it scored 13%, while Claude Sonnet 5.5 (medium) scored 30%. On research-level physics it scored 56%, behind Gemini 3.8 Flash (high) at 65%. On creative tests for 3D and code-made video it beat Gemini 3.8 Flash but trailed Claude Sonnet 5.5.

Its made-up answer rate fell from 61% to 27%. That is mostly because it now attempts fewer questions, 53% instead of 77%. Accuracy rose only from 32% to 34%. In plain terms, it guesses less rather than knowing much more.

What is not known yet

For now it is a proprietary model, available through Thinking Machines' Tinker platform and Tinker Playground. A release of the weights, the files that let others run the model themselves, is planned under an open license called Apache 2.0. The date is unconfirmed. The scores come from Artificial Analysis, so real-world results on your own work may differ.