Definition
Closed models are AI you rent through a vendor’s online service and pay per use; open (open-weight) models are AI you download, run on your own computers, and customize.
At a glance
- Closed (GPT, Claude): no setup, pay per use. Cheapest at low volume, but costs climb fast as you scale.
- Open (Llama, Mistral, DeepSeek): big upfront cost for hardware and engineers, but cheaper per use at high volume.
- Open keeps your data inside your own systems — key for healthcare, finance, and other regulated work.
- Check the license: Apache 2.0 and MIT allow full commercial use; some (Meta’s Llama) add restrictions.
How the money works
Closed bills you per use, so costs grow with volume — one customer-service bot ran ~$50,000/month in API fees.[1] Open flips this to mostly fixed costs (GPUs plus engineers): the same bot self-hosted on Llama cost ~$5,000/month compute plus ~$20,000/month engineering, breaking even in 6-12 months.[1]
Why open is gaining ground
You keep full control and privacy, avoid lock-in, and skip per-use fees.[3] Quality now sits within ~5-10 points of top closed models.[4] An IBM/Morning Consult survey of 2,400+ IT leaders found 51% using open-source AI saw positive ROI, versus 41% who didn’t.[2]
Bottom line
Renting (closed) is cheap and simple to start; owning (open) costs more upfront but wins at high volume with full data control — pick by your volume, privacy needs, and engineering muscle, and often the answer is both.