$34 a month. That is the total cost of running five AI functions across a transportation management center, according to a new framework published on arXiv this week [S1]. The figure sits 97% below the cheapest option using only commercial APIs, and it raises a question every city traffic desk should be asking: which of their AI tasks actually need a premium model, and which can run on something open and free?

The paper, by researchers at Cornell and UC Berkeley, tackles a problem most transport agencies have been solving by gut feel [P2]. Large language and vision-language models are taking on genuine operational roles in traffic management centres, such as detecting unusual activity in camera streams, drafting incident summaries, and creating traveller updates [S1]. Each task has different demands: anomaly detection needs speed, incident reporting needs accuracy, and traveller information needs natural language. Nobody wants to pay GPT-4 prices for every single one.

The maths behind the $34 bill

The authors call their formulation the Foundation Model Deployment Portfolio problem, or FMDP [S1]. At its heart it is a mixed-integer program, a type of optimisation that lowers overall expenses while adhering to quality, latency, and safety limits for every function, all within a shared GPU environment [S1].

They prove the problem is NP-hard, reducing it to the classic 0-1 knapsack problem [S1]. For non-mathematicians: there is no known fast algorithm that always finds the perfect answer. So the authors propose a polynomial-time greedy heuristic, a rule-of-thumb method that runs quickly and produces a good-enough answer rather than a guaranteed best one [S1].

The case study is where the numbers land. Five TMC functions, 19 candidate combinations of model and deployment mode. The heuristic picked a mixed portfolio: four functions routed to open-source APIs, and one function, whose quality floor no open-source model could meet, sent to a closed commercial API [S1]. Total bill: $34 per month [S1].

That is 97% cheaper than the cheapest feasible portfolio built entirely from closed APIs [S1]. The saving comes from the simple insight that most tasks do not need the most expensive hammer in the toolbox.

When buying your own GPU makes sense

The paper also runs a break-even analysis on whether a transport agency should buy its own GPU hardware instead of paying per-query API prices [S1]. The answer: on-premise GPU investment becomes reasonable only above approximately 309 vision queries per hour, or if API prices double from current levels [S1].

Below that threshold, the pay-per-query model wins. Above it, owning the hardware starts to pay for itself. This matters because vision tasks, processing live camera feeds frame by frame, generate far more queries than text tasks. A busy intersection running object detection on every frame could blow past 309 queries per hour in minutes.

The FMDP paper applies the same instinct to deployment: the cheapest AI is the one sized correctly for the job.

What it means

The core idea is that AI deployment is not a single decision. It is a portfolio. A transport agency running five AI functions should not pick one model for all of them, and it should not default to the most expensive option for safety reasons. The FMDP framework forces each function to justify its model choice against concrete quality, latency, and safety thresholds.

The $34 figure is illustrative, not a guarantee. It comes from a specific case study with specific assumptions about model prices, query volumes, and quality requirements [S1]. But the method is the real contribution. Any agency can plug in its own numbers, its own functions, its own quality floors, and get a cost-minimising portfolio that respects safety constraints.

The open-source versus closed-API question gets a clear answer from the data: open-source models handled four of five functions. The one that needed a closed API had a quality floor no open model could meet. That is a snapshot of July 2026. Open models are improving fast. The function that forced a closed API today might not need one in six months.

What it means for business

For a two-person traffic consultancy advising a regional council, the practical takeaway is to audit which AI tasks actually need premium models. The paper's framework suggests most do not. If you are paying commercial API rates for anomaly detection on camera feeds, you may be overspending by a factor of 30.

For a suburban transport agency, the break-even threshold of 309 vision queries per hour is the number to watch. If your camera network generates fewer queries than that, stick with API pricing. If you are processing dense video from multiple intersections in real time, the maths may now favour buying a GPU.

For vendors selling AI services to the transport sector, the paper signals trouble for uniform pricing. If customers can prove that four of five functions run fine on open-source models, the value proposition of an all-in-one commercial package weakens. The customers who run the numbers will cherry-pick.

What we don't know yet

The paper is an arXiv preprint, version one, and has not been through peer review [S1]. The cost figures derive from an illustrative case study, not from a live deployment with real traffic data. The authors' claim that foundation models are "increasingly used" in TMCs is context-setting, not independently verified market data [S1].

The greedy heuristic produces approximate solutions for a problem the authors themselves prove is NP-hard [S1]. How close those approximations sit to the true optimum, across a wider range of functions and model combinations, remains an open question.

The break-even threshold of 309 queries per hour depends on current API prices and GPU costs. If either shifts, the threshold moves. The paper notes that doubling API prices would make on-premise GPUs reasonable at lower query volumes [S1], but does not explore how falling GPU prices or new open-source vision models might change the calculus.

What to watch next: whether any transport agency applies the FMDP framework to real operational data and publishes the results. That would move this from an elegant mathematical exercise to a tested cost-saving tool.

If you found this useful, subscribe for more stories that decode what AI research actually means for the people who have to deploy it.

Sources

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.