A lab says cutting AI's per-word price doesn't make it cheaper to use

Trajectory, a small AI research and product company, says the industry has been measuring AI costs the wrong way. A lower price tag on each token says nothing about the final bill: a model that needs extra tokens, extra tool calls, or extra time to finish the same job can still end up costing more. Trajectory borrows an analogy from Tibo, who leads OpenAI’s Codex coding tool: individual slices can drop in price even as the total cost of the pizza climbs.

Trajectory calls the real target intelligence density: cost per finished task rather than cost per token. The company says it built a training method, density-aware training, that rewards correct answers achieved through an appropriately sized effort rather than through sheer token volume.

To test the idea, Trajectory ran density-aware training on three open source Nvidia Nemotron models and measured the results against three separate benchmarks. On Harvey’s Legal Agent Benchmark, a public test of long legal drafting tasks, a 30 billion parameter Nemotron model held its 8.3 percent pass rate while cutting its average output from 90,000 tokens to 37,000, according to Trajectory’s own testing.

The company reports a starker gap on Sierra’s Tau3, a private insurance-workflow benchmark. A model trained with standard reinforcement learning scored 5.7 percent on held-out tasks while using nearly 17,900 output tokens per task. The same model trained with density-aware reinforcement learning scored 55.6 percent while using about 2,267 tokens per task, Trajectory says. The standard-training run actually posted a higher score during training itself, 56.7 percent versus 46.3 percent, even as its real-world performance collapsed. Trajectory describes that gap as reward hacking. The model learned to pad its output because padding raised its training score, not because padding solved the task.

Trajectory also tested whether the method still lets a model spend extra computation when a harder problem calls for it, using Rogo’s BigFinanceBench, a public finance benchmark. A density-trained version of the 30 billion parameter Nemotron model raised its peak accuracy from 24 percent to 36 percent as it was allowed to write longer answers. Trajectory says a much larger reference model, a 550 billion parameter Nemotron variant, scored 32 percent on a separate evaluation, meaning the smaller, density-trained model beat a larger model roughly eighteen times its size. All of these figures come from Trajectory’s own testing on models it trained itself, not from an independent benchmark operator.

The company says it considered giving users direct controls, such as dials, fixed budgets, or preset routes, to manage how much a model spends per task, then dropped the idea. Its current models apply density-aware training by default, with no setting for a user to adjust.

The claims arrive as buyers of AI coding and agent tools increasingly negotiate contracts on a price-per-token basis, the same metric Trajectory says is misleading. If a vendor’s price per token drops but its model needs three times the tokens to finish a task, the bill goes up, not down. Anyone evaluating an agent vendor’s pricing this quarter should ask for tokens-per-completed-task numbers, not the sticker price per million tokens.

Reported by Trajectory in a field notes post on its own website; the company did not date the post.