The Adoption Shift
By multiple 2026 industry surveys, roughly 89% of enterprises now use open-source AI models somewhere in their stack. DeepSeek V4 matched or exceeded GPT-4o on 7 of 12 standard benchmarks as of April 2026. Open-weight models overall are reported to close 70-90% of the capability gap against closed frontier models at 5-10x lower per-token inference cost.
The Actual Deployment Pattern
Most production systems don't pick a side company-wide — closed models handle general-purpose, unpredictable tasks (open-ended chat, broad reasoning), while fine-tuned open models handle cost-sensitive, high-volume, narrow workflows where a smaller specialized model matches or beats a generic frontier model on that specific task, cheaper.
Why It's a Real Architecture Decision, Not Just Cost
Self-hosting open weights buys control — data never leaves your infrastructure, no rate limits, no exposure to a provider's next pricing change — at the cost of owning the hosting and inference infrastructure yourself. That trade earns its keep for high-volume or compliance-sensitive workloads; for a low-traffic feature, renting a closed API is still simpler and cheaper in practice.
The Practical Takeaway
The useful question per feature isn't 'open or closed' as a blanket policy — it's whether this specific workflow is high-volume and narrow enough that a fine-tuned open model earns back its hosting overhead, or general and low-volume enough that a closed API remains the simpler, cheaper choice.