Google: Gemma 4 31B

Gemma 4 31B is a multimodal model from Google that accepts image, text, and video inputs with a context window of 262,144 tokens and a matching output ceiling. It supports tool use and reasoning, giving it a reasonable footprint for agentic and multi-step workflows. Structured output support is unconfirmed based on available documentation. At $0.12 per million input tokens and $0.35 per million output tokens, pricing sits at the budget end of the multimodal tier. Its blended benchmark score of 56.4 across four benchmarks tells a mixed story: coding reaches 71.7, which is competitive, but agentic performance at 23.8 is notably low, making it a weaker choice for autonomous task pipelines. Teams doing document or video analysis with heavy coding assistance, on a cost-conscious budget, have the most reason to shortlist it; those prioritizing agentic reliability should weigh that score carefully before committing.