The AI Bill Arrives: Tokens, Model Routing and the Cost of Switching Models
For most of the past two years, the question funds asked of an AI tool was simple: can it do the work? Since June, a second question has crept into vendor messaging: what will it cost to run? Research platforms, data vendors and infrastructure providers are all now talking about tokens, routing and model choice. That shift says something about where fund AI is heading, and about who carries the bill.
The pattern
Agents are the trigger. Rogo argues in The New Economics of Agentic AI that pricing is moving from cents per query to dollars per task. It cites a May 2026 Stanford Digital Economy Lab study of coding agents, which found agent tasks used around 1,000 times more tokens than code reasoning or code chat. The same study found that input tokens drive most of the cost, that runs of the same task varied up to 30-fold, and that more tokens did not buy better accuracy. It tested coding, not finance work, but the pattern is a warning for any agent. Rogo’s own agent doubled weekly queries, with sessions now running past 20 minutes.
Vendors are responding with four kinds of cost control.
1. Waste in how models are used. Deal Engine claims in a whitepaper that 92% of Claude token costs are avoidable. Its causes: outputs are not saved, each session rebuilds its context from scratch, and frontier models do routine work. A session with prepared context used about 8,000 tokens against 40,000 from a cold start. In its words, the problem “is how the model is being used”.
2. Routing each task to the right model. Rogo gives a simple example. Building a ten-page deck needs a top model such as Claude Opus 5.5. Tidying up an email does not, so it can go to a model like Gemini 3.6 Flash, which costs eight to ten times less. To make that choice possible, Rogo added Claude Sonnet 5, three tiers of GPT-5.6 and Claude Opus 5 in July alone, each pitched on a different mix of quality, speed and cost. Snowflake went further with dynamic model routing, which picks the model for each task automatically. It claims up to three times better token efficiency on some data work. Blueflame AI makes the same case for matching model to task, though without figures.
3. Leaner context. Hebbia, which processes over 250 billion tokens a month, notes that input can run 10 to 50 times the length of the question asked, and that cached inputs can cut processing cost by up to 90%. Intapp argues that a firm-specific map of business terms cuts both errors and token use, because the model spends less effort working out what the firm means.
4. Models you can swap. Unique pitches AI Conduct as “swappable intelligence”: the workflow stays fixed while the model under each step changes, with cheaper models for simple drafting. Alkymi added Kimi K3, an open model, next to Anthropic, OpenAI and Google, so clients can switch without rebuilding.
The bill is also becoming visible. Marvin Labs prices usage in credits with daily and weekly caps, and raised them sharply after rebuilding its agents to use fewer tokens. Snowflake’s AI Gateway adds spending limits and cost tracking by team and agent. Intapp is even running sessions for private capital leaders titled “The AI bill is growing. The firm isn’t.”
Where vendors disagree
The cheap model is not always cheaper. Hebbia gives the clearest example: a cheaper model that needs 15% of its outputs reviewed by a person can cost more than a dearer one that needs 3%. Retries and review time do not show up on the token bill, but they show up in the P&L. Hebbia also warns that newer, pricier models sometimes do worse on finance tasks than their predecessors.
There is also a split on who chooses. Rogo leaves model choice to the user in a selector. Unique and Snowflake route for them. Each approach shifts the cost decision to a different person.
Why now
Three forces meet. Agents use far more tokens than chat, so costs that were rounding errors are now budget lines. Frontier labs now ship models in tiers, and new versions arrive every few weeks, which turns model choice into a running decision rather than a one-off. And most fund AI vendors run on those same frontier models, so their margins depend on using them well. Cost discipline is as much the vendor’s problem as the client’s, which is why they are talking about it.
What funds should ask now
| Old question | New question |
|---|---|
| Which model does it use? | Which models, for which tasks, and who decides? |
| What is the licence fee? | Is usage capped, metered or bundled, and what happens when agents run longer? |
| How accurate is it? | What does each correct answer cost, once retries and human review are included? |
| Can it use the latest model? | Can we switch models without rebuilding workflows, and can we use our own? |
| Is it secure? | Can we see spend by team, agent and workflow? |
What to watch next
- Pricing models. Expect more vendors to move from seat licences to credits or usage tiers as agent workloads grow.
- Routing claims. Snowflake’s routing is still in preview. Watch for real savings data from fund clients, not vendor test cases.
- Open models. Kimi K3 and DeepSeek now sit beside frontier models on several platforms. Their role in regulated workflows will depend on where they run.
- Review costs. The firms that measure human review time per output will make better model choices than those that only watch tokens.
Qaike tracks how these vendors price and deploy AI in its vendor directory, where you can compare them by function and use case.
This article is based on publicly available information and reflects Qaike’s own analysis and opinions. It is not intended to provide professional advice, endorsement, or a definitive assessment of any vendor or product. While every effort has been made to ensure accuracy, completeness and correctness cannot be guaranteed.
Created by Qaike – Powered by AI.