Your agent bill has an arbitrage in it: a three-tier audit

In August 2026, the Financial Times (FT) reported that 56% of the tokens flowing through Vercel's AI Gateway ran on open-weight models at a significant discount compared to the flagship models. This observation led to the concept of a three-tier audit, which helps businesses understand how much of their agent bill is paying premium prices for work that doesn't necessarily require high-end models.

The audit involves three steps: classifying workloads, pricing each workload under the three tiers, and then reading the results. Workloads can be categorized into three types: tier T1, which consists of bulk or batchable tasks like embeddings, classification, and summarization drafts; tier T2, which involves interactive tasks like user-facing chat and tool-calling agents; and tier T3, which comprises regulated workloads that are pinned to contracted or approved models by policy or compliance.

The pricing of workloads is presented in three tiers: Tier A is the current flagship price, Tier B is the cheapest equivalent flagship, and Tier C is open weights, priced by the gateway-observed ratio for estimation, or by the actual hosting invoice once self-hosting is implemented. The observed ratio of open weights to frontier models is approximately 0.13x, making them roughly 7.8 times cheaper.

To perform the audit, businesses need to price their workloads under all three tiers using the provided script. For example, a workload with 10M input and 5M output tokens per month on Astra would cost $350 per month. The same workload on Sol would cost $70 per month, while on open weights, it would cost approximately $45 per month, after accounting for the gateway ratio.

The transition from Tier A to Tier B saves around 80% of the costs, while moving from Tier B to Tier C is less cost-effective due to the migration effort required.

The audit also highlights a potential trap in pricing: the 272K token limit for Sol pricing, which can result in sudden cost increases when long-context agent loops are involved. Businesses should price their actual request shapes rather than relying solely on monthly totals. The migration process involves four stages: shadowing, canary testing with per-stage hold criteria, and full cutover.

Quality gates, latency budgets, and cost thresholds should be set before the migration begins, and rollbacks should be implemented if any of the predefined criteria are triggered.