Enterprise AI spend rose 108% year-over-year in 2026, to an average of $1.2 million per organization, with 78% of IT leaders reporting unbudgeted charges, even as the price of a single token kept falling (BERI). That gap is the story: model intelligence and cost control now hinge on a different question than “which model scores highest on the leaderboard.” The question that actually determines a budget is cost per completed task.
Per-token prices really have collapsed. GPT-4o’s input price fell from $5.00 to $2.50 per million tokens, and efficient models like o4 Mini now price input at $0.55 per million tokens (Iternal.ai). Cheapest-tier models, including DeepSeek and Gemini Flash-class models, now run under $0.50 per million input tokens, versus $15 to $75 per million output tokens for frontier models, a spread of more than 100x across the market (AI Pricing Guru).
Why Is Enterprise AI Spend Still Rising If Prices Are Falling?
Usage is growing faster than prices are falling. Inference now represents roughly 85% of enterprise AI budgets, and 60% of AI projects exceed their cost estimates by 30 to 50% (BuildMVPFast). Agentic workflows, where a model plans, calls tools, and checks its own work across multiple steps, consume five to thirty times more tokens per completed task than a single chat exchange. Multiply that by how many more agentic workflows enterprises are now running, and a falling per-token price still adds up to a bigger total bill.
Cost Per Task Is the Metric That Actually Matters
Artificial Analysis’s Intelligence Index v4.1 now tracks cost-per-task directly rather than just cost-per-token. In one comparison, GPT-5.6 Sol scores just one point below Claude Fable 5 on overall intelligence, but at roughly one-third the cost, about $1.04 per task at maximum reasoning effort (Artificial Analysis). That distinction matters because raw benchmark leadership only tells a buyer which model is smartest in isolation. It doesn’t say what a specific workload will actually cost to run at scale, and those two numbers can point to different vendors.
The Real Levers for Cost Control
Three practical levers are doing most of the work for teams actually managing this. Prompt caching, storing the parts of a prompt that repeat across calls so the model doesn’t reprocess them each time, cuts cached input costs by 75 to 90% on providers that support it. Routing simple sub-tasks to smaller, cheaper models instead of a single flagship model for every step is the second, and it’s covered in more depth in this series’ piece on model routing. The third, and the simplest to act on immediately, is just matching model choice to task complexity rather than defaulting to the most capable (and most expensive) option for every request.
This is also where Model Flexibility becomes a practical design choice rather than a theoretical one. Commercient’s AI agents aren’t locked to a single provider: a team can point one agent at a lower-cost model for routine data lookups and another at a stronger model for judgment calls on a sales opportunity, and switch either one the moment a better or cheaper option ships, without rebuilding agent roles, permissions, or knowledge bases. Usage is billed by token, so switching models doesn’t come with a surprise invoice.
See how Commercient’s Model Flexibility lets your AI agents run on the model that fits the task, and the budget. Book a free demo.
Frequently Asked Questions
What does “cost per task” mean in AI pricing? It measures the total cost to complete a defined unit of work, such as answering a support ticket or drafting a summary, rather than the price of a single token. Comparing models at the task level accounts for how many tokens a model needs to reach a correct answer, not just what each token costs.
Why are AI bills rising even though token prices are falling? Usage is growing faster than prices are falling. Agentic workflows that involve multiple steps and tool calls consume many more tokens per completed task than a single chat exchange, and enterprises are running more of these workflows at greater scale.
What is prompt caching and how does it cut costs? Prompt caching stores the parts of a prompt that repeat across calls, such as system instructions or reference documents, so the model doesn’t reprocess them each time. Providers that support it price cached input tokens 75 to 90% lower than fresh input.
Should enterprises always use the most capable model available? Not for every task. Using a frontier model for straightforward, high-volume work often costs far more than the task requires, which is why many enterprises are matching model choice to task complexity instead of defaulting to one model for everything.
How does Commercient help control AI model costs? Commercient’s Model Flexibility capability lets agents point at any commercial, cloud, or self-hosted model and switch at any time, with usage billed transparently by token, so teams aren’t locked into paying frontier prices for routine work.
