Inkling 2 Small scores 31 on the Artificial Analysis index and costs $1.18 per task

Thinking Machines has released Inkling 2 Small, a model that scored 31 on the composite index from the independent lab Artificial Analysis. The index combines dozens of tests: knowledge, coding, expert-level questions and agent work. A score of 31 is 18 points higher than the previous Inkling Small and equal to Google DeepMind's Gemini 3.

8 Flash (high). The second number in the headline, $1.18, is the average cost of one task in the index: what you would pay if you gave the model a typical assignment and waited for the finished result.

For comparison, the same task costs $7.77 on Gemini 3.8 Flash (high) and $10.

02 on Claude Sonnet 5.5 (medium). Put simply: if you are choosing a model for an agent that edits code, compiles reports or fills in spreadsheets on its own, a cheap token does not mean cheap work, and this is a case where per-task pricing favors the new model.

For now, Inkling 2 Small is closed. It is available only on the Thinking Machines Tinker platform and in Tinker Playground. The company plans to release the weights under the Apache 2.0 license later, but the date is not confirmed. If it keeps that promise, this would be the strongest open model built outside China by current rankings. It also outscored Xiaomi's MiMo-V2.6-Pro. It is too early to build an open version into your plans.

Where the model is strong and where it lags

This is a reasoning model with 594 billion parameters, of which 17 billion are active for each token. It accepts text and images and responds with text. The biggest gains came in agentic tasks, where the model has to carry the work through to the end on its own:

• AutomationBench-AA, workflow automation: from 5% to 46%.
• FrontierCode 2.0, hard programming tasks: from 5% to 25%; Claude Sonnet 5.5 (medium) scores 11%.
• HLE-Diamond, expert-level questions: from 12% to 29%.
• AA-Briefcase, long office workflows with spreadsheets, presentations and memos: 1527 Elo versus 912 for the previous version, 1439 for Sonnet 5.5 and 1203 for Gemini 3.8 Flash. The pass rate on checks rose from 20% to 50%.

In building interactive 3D scenes (AA-3D) and generating video through code (Code2Video-AA), the model beat Gemini 3.8 Flash, but Claude Sonnet 5.5 stays ahead in both tests. There are weak spots too: in Terminal-Bench 4.0, where an agent works in a command line, Inkling 2 Small scores 13% versus 30% for Sonnet 5.5. In research-level physics (CritPt v2), it scores 56% versus 65% for Gemini 3.8 Flash (high). So for complex edits in a repository the result is better, while for terminal work it is noticeably worse.

Fewer fabrications, because it stays silent more often

In the AA-Omniscience v1.1 test, the hallucination rate fell from 61% to 27%. Accuracy, however, rose only from 32% to 34%. The main reason lies elsewhere: the model now attempts to answer 53% of questions instead of 77%. When it is unsure, it declines more often, so it states falsehoods less often. The overall score rose from −9 to 16, but Gemini 3.8 Flash (high) at 23 and Sonnet 5.5 (medium) at 19 are still higher. If a wrong answer costs more than a refusal in your product, as in support or legal reference, this behavior is useful. If a chatbot must always answer, you will need to handle refusals separately.

Why expensive thinking still comes out cheap

On average, Inkling 2 Small spends about 150K output tokens per index task. The previous version used about 37K, Sonnet 5.5 (medium) about 31K, and Gemini 3.8 Flash (high) about 99K. As a result, the cost per task rose almost 6-fold, from $0.20 to $1.18.

The rates, however, are low: $0.30 per million input tokens, $1.00 per million output tokens and $0.03 per million cached input tokens. That is why, even with enormous token consumption, the model is cheaper than neighbors with a similar score.

What to do with this

Teams building agents for office automation, reporting and interface generation should run their own scenarios on Inkling 2 Small and compare the price of a completed task, not the price of a token. Long reasoning raises latency, so the model may be slow for interactive chats. The figures come from Artificial Analysis benchmarks and reflect its set of tasks; results on your data may differ. Open weights will interest those who want to deploy the model themselves, but for now that is a promise without a date.