Open models can do 90 percent of agentic work at a twentieth of the cost of a frontier LLM: Gartner

Gartner analysts project that by 2030, open-weight models will execute 90% of enterprise operational agent-workflow steps. They say open models can cost one-quarter as much as the cheapest closed model at the same score, or as little as one-twentieth of using a frontier model on every step; allowing for verbosity and retries, savings versus frontier models are closer to five or six times per task. The analysts argue that open weights preserve enterprise control, while agent harnesses that route work, enforce guardrails and manage fleets are a key source of product value.

Analysis

The key exposure is value capture, not model share. If routine inference commoditizes, closed-model providers risk lower pricing power and weaker returns on infrastructure unless agent volumes expand enough to offset falling revenue per task. The likely beneficiaries are orchestration, evaluation, security and workflow-control layers that can route tasks across models; enterprises with proprietary workflow data may also retain more strategic value by keeping routine execution in-house. Hyperscalers are mixed: lower unit costs can pressure premium API economics, while increased inference volume and private deployments can support cloud and accelerator demand.

The contrarian point is that Gartner’s 90% refers to workflow steps, not spend, compute, or economic value. A small number of difficult steps could absorb disproportionate inference cost and remain frontier-model dependent. Moreover, per-task model pricing is not total cost of ownership: retries, hosting, security, maintenance and governance could narrow open-weight savings. The 2030 adoption claim is therefore a directional risk to model rents, not proof of near-term revenue displacement.

Near term, this is more a valuation and product-positioning catalyst than an earnings event. Over 1–3 months, watch enterprise disclosures on model-routing deployments and inference cost per completed workflow. Over 6–18 months, the thesis strengthens if agent usage scales while model/API pricing falls and orchestration vendors show monetizable customer adoption. It weakens if frontier-model providers sustain premium pricing or enterprises avoid self-hosting due to operational and security burdens.

Key Decisions for Investors

• Avoid treating this as an immediate short signal on all AI infrastructure: cheaper inference may expand total usage, and the article provides no company-specific financial evidence. Reassess model-provider exposure against realized pricing, utilization and inference revenue disclosures.
• Prefer a relative-value watchlist: favor enterprise orchestration, agent-management, evaluation and security platforms over undifferentiated model/API exposure, but require evidence of paid adoption and retention before initiating longs. The thesis is falsified if these tools remain bundled or fail to convert workflow control into pricing power.
• Track cloud and accelerator exposure as a two-sided trade rather than assuming open weights mean lower compute demand. A sustained decline in cloud inference growth alongside falling API prices would support reducing exposure; rising agent workloads and private deployment demand would argue against it.
• Monitor total cost per successfully completed workflow—not headline token cost—including retries, hosting, governance and maintenance. If open-weight deployments fail to deliver durable TCO savings, the expected shift in enterprise economics and model bargaining power is overstated.