The cheapest-looking setting produced the most expensive result.
On CursorBench 4.0, Google's new Gemini 4 Carbon model scored 60.4% in high reasoning mode at an average cost of $9.22 per task. Medium scored 60.1% for $8.49. Low scored 58.6% and cost $11.89.
Turning reasoning down did not merely save less than expected. In this benchmark, it cost 29% more than high while doing slightly worse. The low setting used more reported tokens and took 53 more agent steps: 42,663 tokens and 124 steps, compared with 39,960 tokens and 71 steps for high.
That is a useful little reversal. It says something larger about how we price reasoning models: a cheaper decision is not always the one that thinks less at each step.
The meter can run while the model recovers
Google released Gemini 4 Carbon to paying developers on November 10, describing it as the heavier model for long coding jobs, scientific work and extended verification. Its API price is currently $2 per million input tokens and $10 per million output tokens. Those rates double on December 1. Output billing includes the model's hidden thinking tokens, and tool instructions count too.
The price sheet therefore tells you what the meter charges, but not how long the meter will run. A low reasoning setting may spend less deliberation on one decision, choose a weaker path, discover the mistake later, call another tool and try again. High reasoning can be expensive per turn yet cheaper across the whole job if it avoids enough recovery work.
CursorBench's rows fit that explanation, but they do not prove it. The public table gives aggregate tokens, steps, scores and cost. It does not provide a neat causal autopsy for why Carbon Low took 124 steps. The honest conclusion is narrower: on this workload and harness, the low setting was not a budget cap. It cost more, used more reported tokens and took more steps.
A token is a unit of usage, not value
Token prices are easy to compare because every vendor prints them. Completed tasks are harder. You need a definition of success, a realistic workload, a harness that lets the model use tools, and a decision about how to count partial failures. All of those choices can move the answer.
Still, cost per task is closer to the question a buyer usually has. Nobody wants 100,000 attractively priced tokens as a decorative object. They want the migration completed, the bug fixed or the research checked. If the cheap model needs three attempts and the expensive model needs one, the token menu has concealed the relevant arithmetic.
Artificial Analysis reached a similar, though not identical, result on its broader Intelligence Index. Using Carbon's standard December pricing, it calculated $7.76 per task: less than Claude Opus 5.5 at $10.04, but more than GPT-6 Astra at $6.25. Carbon led that index overall, yet it did not lead every coding or knowledge-work category. “Best” changes when the workload changes, which is inconvenient but considerably more useful than pretending one leaderboard row is a law of nature.
The labels describe effort, not a bill
Google's model card says Carbon's thinking levels are configurable to balance quality, latency and cost. That is sensible wording. A balance is not a guarantee, and “low” describes how the model is asked to reason, not the maximum amount your workflow can spend.
One benchmark cannot tell you which model will be cheapest for your codebase. Cursor also warns that results vary and small score differences may not be statistically meaningful. The three Carbon scores are close enough that the ranking deserves restraint. The cost reversal is much larger, but it still belongs to CursorBench's tasks, tools and current prices.
The practical test is dull and therefore good: run representative jobs, count failures and retries, include tool charges, then divide total spend by accepted results. Also keep a hard budget outside the model, because a dropdown labeled “low” has no fiduciary duty.
Gemini 4 Carbon may be unusually capable. The more durable lesson from its launch is that intelligence has made the invoice harder to read. The token is still useful for billing. For choosing a model, the completed job is the unit that matters.