Gartner argues open-weight models can handle about 90% of routine enterprise agentic workflow steps at a fraction of the cost of frontier or closed models, making them highly cost-effective for enterprises.

Real-world open models can be more verbose and require more retries, which reduces savings to roughly five to six times cheaper per task than frontier models.

Analysts Mark Petty and Raymond Paquet estimate typical enterprise workflows involve hundreds of routine tasks—classification, tool calling, data passing, validation, reconciliation—open-weight models can handle these at parity for roughly a quarter the price of the cheapest closed model and as little as a twentieth the cost of running every step on a frontier model.

Harnesses, the systems that route tasks to models and manage fleets, are where most value lies in AI products, unlocking cost savings, guardrails, and efficient agent management.

The authors frame harnesses as the source of value since they route tasks to appropriate models, enforce guardrails, and manage fleets; open-weight models preserve enterprise control by keeping data perimeters intact.

Overall takeaway: enterprises should balance frontier models for R&D with open-weight models for routine, high-volume tasks, using harnesses to optimize routing and governance.

Analysts advise planning for a split now: build robust harnesses to route routine steps to open weights and reserve frontier APIs for rare steps requiring advanced reasoning, as the workflow layer controls capture most value.

Even with cost considerations, the main value lies in the orchestration layer: open-weight models help keep execution inside enterprise boundaries, preserving control and reducing leakage of workflow history outside the perimeter.

Strategic shift emphasizes domain-specific open-weight models for highly repetitive, high-volume steps, where stability and repetition trump novelty.

Open-weight models offer control and perimeter security advantages since routine steps do not exit the enterprise boundary, unlike frontier-driven processing.

Frontier models will still be used for research and development as a capital expenditure, with frontier reasoning handling heavy lifting before delegating routine tasks to open-weight models as operational expenditure.

Enterprises will treat most AI models as commodities and focus on the harness or orchestration layer that assigns tasks to the appropriate model, manages fleets, guardrails, and user experience.