Open-weight AI models can perform 90% of the routine steps in enterprise agent workflows at a twentieth of the cost of running a frontier large language model (LLM) for every step, according to Gartner analysts.
They argue that most of what enterprise AI agents actually do, classification, tool calls, data transfer, validation and reconciliation, is routine rather than reasoning-intensive, meaning cheaper open models can match frontier performance on those steps.
Gartner puts the raw saving at one-quarter of the cost of the cheapest comparable closed model, rising to one-twentieth when measured against using a frontier model for every step throughout a workflow.
Verbosity and retry rates erode some of that saving in practice, cutting it to roughly 5 to 6 times cheaper rather than 20 times.
Gartner's planning assumption is that open-weight models will handle 90% of enterprise operational agent steps by 2030.
Building cutting-edge AI is eye-wateringly expensive, as Anthropic's updated listing documents show.
The company has committed to spend nearly $900 billion on computing power through to 2045.
Computing and infrastructure costs reached about $10 billion in the third quarter alone.
That helped push it to a $567 million operating loss, even though revenue climbed to around $17 billion.
Read more
