AI is getting cheaper at the same time that companies are spending more on it. That sounds contradictory, but it is one of the most important things happening in enterprise AI right now.

The price of processing a million tokens keeps falling, especially as open-weight models improve and compete aggressively on cost. At the same time, enterprises are moving from small pilots to always-on copilots, automated workflows and multi-step agents. Cheaper units make it easier to use more units.

For insurance and financial services, I think this changes the AI conversation in a useful way. The question is no longer which single model should win. The better question is which model should handle which kind of work, at what reliability level and under what governance rules.

Tokenomics Is Becoming a Real Operating Discipline

I use tokenomics to describe the economics of AI at the unit level. Every prompt, retrieved document, tool response and generated answer consumes tokens. Once AI moves into production, those tokens stop being an abstract engineering metric and become a recurring operating cost.

Pricing is also becoming more complicated than a simple monthly subscription. Many enterprise deployments now combine usage-based pricing, committed capacity, caching discounts and different rates for input and output. A team can have a cheap model and still end up with an expensive system if prompts are bloated, agents loop unnecessarily or every task is routed to the most capable model by default.

That is why lower token prices do not remove the need for cost governance. They actually make governance more important because falling prices encourage more experimentation and more volume.

The Open-Weight Shift Is Bigger Than One Model

The most interesting part of the current market is that the pressure is not coming from one challenger. It is coming from an ecosystem of open-weight models that keeps improving. Families such as DeepSeek, Qwen, Kimi, GLM and MiniMax have made it harder to assume that the best enterprise option must always come from a closed frontier provider.

That pressure is visible in pricing. In May 2026, DeepSeek made a 75 percent reduction on its flagship V4-Pro pricing permanent [1]. The exact leaderboards will continue to move, but the broader direction is clear: strong models are becoming available at much lower unit costs than many teams were planning around a year ago.

For enterprises, this creates something closer to a portfolio market. Instead of asking one model to do everything, teams can route work based on task difficulty, latency needs, data sensitivity and the cost of being wrong.

The Cheapest Token Can Still Be the Most Expensive Decision

This is where regulated industries need a different lens from consumer experimentation. Insurance and financial-services workflows often involve underwriting, claims, customer communications, compliance reviews and other decisions where a small quality gap can create a large downstream cost.

Reliability becomes especially important in agentic workflows. If a model succeeds 95 percent of the time at each step, a five-step workflow succeeds end to end only about 77 percent of the time. Each step may look strong in isolation, yet the combined workflow can still fail often enough to create rework, manual review and customer friction.

That is why I do not think the right strategy is to route everything to the cheapest model. A cheaper model may be perfect for summarization, classification, drafting or high-volume internal tasks. A higher-cost model may still be worth the premium when a workflow is complex, regulated or difficult to recover from when it fails.

A Better Model Strategy Looks Like Routing

The enterprise AI stack is starting to look less like a single vendor decision and more like a routing problem. The goal is to match model capability to business risk. I call this approach a risk-tiered model routing framework: work is classified by business risk first, and model choice follows from that classification rather than the other way around.

Routine work can often go to smaller or open-weight models. Medium-risk tasks can use stronger models with additional checks. High-impact decisions can be routed to the most reliable model available and paired with human review, deterministic rules or independent verification.

This is not only about saving money. Routing gives teams a way to make cost and reliability explicit. It forces product leaders to define which tasks actually require frontier-level capability instead of paying the reliability premium everywhere.

In my own work leading AI product strategy for large financial-services and insurance organizations, this kind of tiering plays out in practice. Routing high-volume, low-risk tasks such as document summarization, intake classification and internal drafting to smaller or open-weight models has meaningfully reduced inference cost without a measurable drop in output quality, while tasks tied to underwriting judgment, compliance communications or customer-facing decisions have continued to justify a reliability premium, paired with human review. The savings from the first category are what fund the ability to be conservative in the second.

Open Weights Also Change Vendor Risk

Cost is only one reason open-weight models matter. They can also change how organizations think about continuity and control. A model that can be deployed in more than one environment gives a team options if a hosted service becomes unavailable, changes terms or becomes difficult to use in a particular jurisdiction.

That became more than a theoretical concern in 2026, when a U.S. export-control action temporarily suspended access to certain frontier AI models worldwide [2]. Whatever your view of those policies, the engineering lesson is straightforward: dependency on a single model endpoint is now a business-continuity question.

For regulated companies, model provenance belongs in the same conversation. Teams need to know where a model came from, what license applies, how it was trained or fine-tuned when that information is available, where inference runs and who is responsible for monitoring changes. Open does not automatically mean low risk, just as closed does not automatically mean safe.

FinOps for AI Needs to Start Before the Bill Hurts

Traditional cloud FinOps taught teams to track compute, storage and network consumption. AI needs a similar discipline, but token count alone is not enough.

I would track cost by use case, model, workflow and outcome. A workflow that costs twice as much but cuts manual review dramatically may be the better investment. Another workflow may look cheap per call but become expensive because it runs millions of times or repeatedly retries failed steps.

Teams should also watch prompt size, retrieved context, output length, cache hit rates and the number of model calls inside one user action. Those are the places where an apparently small design decision can quietly multiply production cost.

Falling Prices Will Probably Increase Total Spend

There is a larger paradox here. AI infrastructure spending continues to rise even while inference prices fall. That makes sense if cheaper AI unlocks more use cases, more users and more automation.

A team that once used AI for one chatbot may now use it for document processing, customer support, coding, rese