# Gemini 3 Flash on the Cotera agent benchmark

> Cheap on the easy ones, ate $0.34 on AirPods complaints. Flash that occasionally remembers it's Pro.

Source: https://cotera.co/benchmarks/models/gemini-3-flash

---

Google's Gemini 3 Flash — the cost-tier sibling of Gemini 3 Pro. Optimized for latency and price; reasoning depth is supposed to be the tradeoff. Hosted on Vertex.

**Score:** 5/5 benchmarks passed · **Cost across the matrix:** $0.758

## Strengths
- Three tool calls and $0.032 on the Crunchbase brief — the cheapest pass on Sales in the matrix until GPT-5.6 Terra undercut it.
- Five for five, no JSON drama. The native Gemini structured-output mode handles the schema constraint without prompting acrobatics.
- Total spend was $0.76 — still well below every Claude model and both GPT-5.5 and Gemini 3.5 Flash.

## Watch-outs
- Pricing wobbled. The CX benchmark hit $0.34 (13 tool calls) — 10x more than Crunchbase. Flash's instinct to over-call when the task is open-ended is real.
- Marketing ($0.22) cost more than 5x Mistral's run. Reddit threads with long comment chains burn a lot of input tokens at Flash's per-token rate.
- We didn't see any reasoning failures, but the model is documented to drop schema fields under high-context load. Worth verifying before production.