April brought another strong signal that open models have grown up. Google released Gemma 4, describing it as "byte for byte the most capable open model" and built for reasoning and agentic workflows, and noted that developers have downloaded Gemma models more than 500 million times. For teams building products, the question is no longer whether open models are good enough, but when they are the better choice.
When an Open Model Makes Sense
Open-weight models shine when you need control. If customer data cannot leave your infrastructure, if you want predictable costs at high volume, or if you need to fine-tune behaviour for a narrow task, running your own model can beat paying per request. They also protect you from a single vendor changing prices or retiring a model you depend on.
When a Hosted API Is Still Better
Hosted frontier models remain the fastest way to ship. You avoid GPU capacity planning, scaling and model operations, and you get the newest capabilities the day they launch. For early products, low or spiky traffic, and tasks that need the very best reasoning, an API is usually the cheaper and simpler path.
A Practical Way to Decide
We recommend starting with a hosted model to prove the feature is valuable, while keeping the model behind a thin interface in your code. Measure real usage, cost per task and quality on your own examples. Once volume is predictable, test an open model against the same examples. If it matches quality for the task, moving that workload in-house can cut costs and give you more control, without rewriting the product.
The Takeaway
Treat the model as a replaceable component, not the foundation. Teams that design for that flexibility can take advantage of releases like Gemma 4 without being locked into any single provider.