One Hundred Dollars: 10 Tasks on Opus 5.5 Max or 2,500 on GPT-6 Luna

Model price lists have long since turned into a puzzle. The price per million tokens looks straightforward until you discover that one model “thinks” ten times longer than another and that reasoning mode can change the bill by an order of magnitude. Artificial Analysis proposed a more honest metric: how many Intelligence Index tasks can be completed on a fixed $100 budget.

What the Numbers Showed

The task set is mixed and includes different archetypes, ranging from reasoning to practical scenarios. The results are as follows:

• Opus 5.5 at the medium reasoning level: 40 tasks.
• Opus 5.5 at the maximum level: 10 tasks.
• DeepSeek V4.1 Flash 1017 at maximum: 166 tasks.
• GPT-6 Luna at the medium level: 2,500 tasks.

That works out to approximately $2.50 per task for Opus 5.5 medium, about $10 for the same model at max, roughly $0.60 for DeepSeek, and about 4 cents for Luna. These are our estimates based on the post's data, not official prices. But the order of magnitude is clear: the gap between the most expensive and cheapest options is 250-fold.

The Main Lesson: Mode Matters More Than the Model

Look at the Opus figures. It is the same model, but switching from medium to max makes every task four times more expensive. “Thinking” tokens cost money, and the more demanding the setting, the more of them are consumed. Many teams select the maximum by default “to be on the safe side” and then wonder why the bill is so high.

Imagine a support operation where AI processes 20,000 requests per month. On Luna, that would cost around $800 if calculated using the benchmark's per-task price. Opus 5.5 medium would cost about $50,000, while max would come in at around $200,000. Real support requests may be easier or harder than the index tasks, so these figures are illustrative. But their scale shows why model and mode selection cannot be based on intuition alone.

What to Do About It

First, calculate the cost per solved task, not per token. Run 50–100 of your real cases through two or three models and modes, then record the price and success rate. The cost of a successful answer, rather than the cost of a call, is what determines the economics.

Second, build routing. Send simple requests—classification, field extraction, and short answers—to inexpensive models such as Luna or DeepSeek Flash. Escalate difficult cases where the cheaper model fails to Opus 5.5, and enable high reasoning only there. This setup often cuts the bill severalfold with almost no loss of quality.

Third, keep the reasoning mode in configuration rather than in code. This lets you adjust it by task type and according to measurement results.

Limitations

The post omits the most important detail: how many of these tasks each model solved correctly. A $100 budget shows the price of attempts, not quality. A cheap model that failed half the tasks may actually cost more than it appears. Moreover, the Intelligence Index task mix does not match your workload: the figures may differ substantially for agentic scenarios with long chains of calls. The data have so far been presented only in a tweet, without a detailed methodology, so treat them as a reference point rather than a verdict.

The conclusion is simple. The era when it was enough to choose the “smartest” model is over. Today, the winner is the team that can match the model and reasoning level to each task and knows the cost of one correct answer.