Research review / models

Gemini 4 Neon Makes Thinking Levels a Task Cost Decision

Google's November 18 release adds four thinking levels. Its efficiency claims make completed-task cost a useful evaluation target, provided retries and incorrect answers remain in the accounting.

By Sandeep Belgavi · Published 2026-11-19

Google announced Gemini 4 Neon on November 18, 2026, including availability through the paid Gemini API. The release offers four thinking levels, making the reasoning setting an explicit choice for each workload.

What Google's release establishes

The announcement lists a one-million-token context window and prices of $3 per million input tokens, $0.15 for cached input and $15 for output. Google reports that the default high thinking setting costs 32% less per task than Carbon in its comparison. These are vendor-reported specifications and results; this digest has not reproduced the comparison.

Count the work needed to get an accepted result

Illustrative calculation: assume an uncached request contains 100,000 input tokens and produces 2,000 output tokens. At those rates, the token charges are $0.30 plus $0.03, totaling $0.33. This excludes tool costs and other charges and is not an observed bill.

For a workload decision, an inexpensive incorrect response still creates work. A useful comparison should include every attempt required to obtain an output that passes the declared checks, while recording tasks that never pass separately.

A proposed test of the reasoning setting

Original exercise, not executed: generate 20 small parser fixtures with known outputs and deterministic assertions. Run the same fixtures at each thinking level with identical tools and retry limits. Record assertion passes, total tokens, retries and elapsed time.

If one incorrect answer is followed by a successful retry, count one completed task and the cost of both calls. Freeze the fixtures before comparing settings. This would test a local cost hypothesis without treating a vendor leaderboard as a prediction for the parser workload.

Sources accessed November 21, 2026. Views are my own and do not represent any employer, past or present.

Sources and evidence

Compare reasoning settings using accepted outcomes and the cost of all attempts, including retries and tasks that never pass.