Inkling 2 Small Uses 150K Tokens per Task—4x More Than Its Predecessor

When choosing a model for production, the first thing many people check is the price per million tokens. It is a convenient but misleading metric. New data from Artificial Analysis clearly shows why.

What the measurements showed

When running the Intelligence Index, Inkling 2 Small generates around 150,000 output tokens per task on average. The previous Inkling Small used around 37,000, meaning consumption has increased roughly fourfold. For comparison, Gemini 3.8 Flash in high mode produces about 99,000, while Claude Sonnet 5.5 in medium mode generates around 31,000.

Importantly, these are output tokens—that is, the reasoning and the answer. They typically cost more than input tokens and also determine how long users have to wait.

Why this matters in practice

Imagine you are building an agent that processes support tickets. Model A costs half as much per token but uses four times as many tokens for each ticket. It ultimately costs twice as much and responds noticeably more slowly. Users do not see the price list; they see the waiting time and the bill at the end of the month.

The same applies to limits: long reasoning chains hit the context window and provider rate limits more quickly. Agentic workflows, in which a model is called dozens of times in succession, multiply the effect.

How to deal with it

First, compare models by the cost of a completed task, not by token price. Artificial Analysis publishes precisely these kinds of breakdowns, although the picture may differ with your own data.

Second, check whether you can cap the reasoning budget. If the model offers modes or parameters controlling reasoning depth, run your tasks at several levels and identify where quality stops improving.

Third, monitor both the average and p95 output-token counts. A sharp jump after a model version update is easy to miss, but it will show up immediately in the budget.

Limitations

The post does not make clear how much the Intelligence Index score improved alongside the increase in token consumption. If the quality gain is substantial, fourfold consumption may be justified for complex tasks. If not, the model is simply more verbose. This is also a benchmark average, and your workloads may behave differently. The post itself received few views and provides only the figures without a methodology, so the full picture should be checked on the Artificial Analysis website.

Conclusion

Price per token is no longer an honest metric on its own. Token efficiency should be evaluated alongside quality and latency, especially when building agents or high-volume pipelines. Before migrating to Inkling 2 Small, run your own task set and compare its total cost and response time with Sonnet 5.5 and Gemini 3.8 Flash.