Gemini 3.6 Flash: the Thinking Dial That Moves Cost 30x (Measured)

Short summary

Gemini 3.6 Flash bills reasoning tokens at the full output rate, and the reasoning_effort dial controls how many it spends—swinging per-call cost up to 30x on the same task. Setting it to 'minimal' cut costs 91–97% on extraction, lookup, and formatting tasks while preserving accuracy, but broke multi-step math entirely. The 1M context window and prompt caching floor matched Google's published specs exactly. The practical takeaway: set reasoning_effort per route, not globally, and verify accuracy on reasoning-heavy workloads before shipping.

• •reasoning_effort 'minimal' cuts cost 91–97% but breaks multi-step reasoning tasks
• •Reasoning tokens bill at full $7.50/M output rate and can outnumber answer tokens 30:1
• •1M context and prompt caching match Google's published specs exactly

Generated with AI, which can make mistakes.