Google's Gemini 3 Pro Costs $12 per Million Tokens—and It's Cheaper Than GPT-5

The price sheet is the story

Google published Gemini 3 Pro pricing on October 24: $2 per million input tokens and $12 per million output tokens. That is $4 below OpenAI’s GPT-5 output rate of $16 per million, per OpenAI’s published pricing page.

Input is the same $2 as GPT-5. The delta is entirely on generation, which is where agent workloads spend most of their budget.

Google is not discounting a weaker model. On the published model card, Gemini 3 Pro leads GPQA Diamond at 91.2% and SWE-bench Verified at 76.4%. GPT-5 reports 88.4% and 74.9% on the same evaluations.

What the benchmarks don’t cover

Benchmark deltas of two to three points rarely survive contact with a real workload. GPQA Diamond is a graduate-level science set; SWE-bench Verified is curated GitHub issues. Neither tells you how a model behaves on your schema, your retries, or your tool contracts.

The more useful number is cost per completed task. A model that scores lower but needs fewer repair loops can win on total spend. Google’s own pricing page lists context caching at $0.50 per million tokens for Gemini 3 Pro, which changes the math for long-running agents that reuse system prompts.

The cheaper tier is the sleeper

Gemini 3 Flash ships at $0.30 input and $2.50 output per million tokens. Google positions it for high-volume classification and routing. At those rates, a Flash-based router in front of Pro costs less than a rounding error on most agent bills.

The pattern echoes what smaller open models have been doing all year: route cheap, escalate rarely. The difference is that Google now sells both ends of that ladder from one API.

What to check before migrating

Rate limits and regional availability differ by tier. Gemini 3 Pro is listed in the Gemini API docs with a 1M-token context window, but throughput quotas are per-project and not published in the pricing table.

Run your own eval before switching. Take twenty real prompts from production, run them through both models, and compare tokens billed—not tokens generated. Output length alone can erase a $4 price gap.

Google’s pricing page is the source of truth, and it changed twice last quarter. Pin the numbers in your cost model with a date.