DeepSeek Slashed Prices 75%. AI Costs Are Still Rising.

Key points

• DeepSeek cut the price of its V4-Pro large language model by 75% in mid-2025, but many business customers say their AI bills are still climbing.
• A single user request handled by an AI agent, software that carries out multi-step tasks on its own, can trigger 35,000 or more billable units of text, called tokens, where a basic chatbot would trigger a few hundred.
• One realistic enterprise query on a leading model costs between $0.10 and $0.40; at one million queries a month, that is a six-figure line item before any other expenses.
• Several enterprise software vendors are privately reporting negative gross margins on their heaviest AI users, meaning those customers cost more to serve than they pay.
• OpenAI recently offered every startup in the Y Combinator accelerator $2 million in free API credits, a sign of how expensive it now is to build an AI-native product from scratch.

DeepSeek, the Chinese AI lab, cut the cost of its V4-Pro model by 75% earlier this year. For companies building AI products, that should've felt like a pay rise. It hasn't.

The reason sits in how modern AI systems are actually built.

A basic chatbot is simple: one question, one answer. An AI agent is different. It breaks a single user request into many smaller steps, planning, searching for information, using tools, checking its own work, then writing a final answer. The user sees one response. Under the hood, the vendor pays for every step.

That gap is now called the 100x problem. A query costing a few cents on a chatbot can cost dollars on an agent system.

Here's a concrete example. A user types: "What did our top customer ask about last week?" A typical agent touches seven separate billed operations to answer it: reading the user's message, loading its system prompt and tool definitions, pulling in relevant records, choosing a tool, running it, summarising the results, deciding whether to follow up. Together those steps bill roughly 35,000 tokens. Tokens are the small chunks of text that AI models count and charge for.

At current prices on a frontier model, that one query costs $0.10 to $0.40. Across one million queries a month, a typical target for any business software feature, the bill lands in the six figures.

Why does this matter to ordinary software customers?

It matters because the standard pricing model for business software breaks under this pressure. Most AI software today sells on a per-seat, per-month basis: a flat fee for each employee who uses the tool. That works fine when costs per user stay predictable. Token amplification destroys the prediction.

A power user running 50 to 100 agent tasks a day on a $40-per-month plan can cost the vendor more in AI processing than that plan collects in revenue. The vendor loses money on its best customers. That paradox, as VentureBeat reported this week, is now showing up in real company accounts.

Salesforce's Agentforce product is the most public example. Bloomberg documented a widening gap between the marketing demos and what's actually shipping to customers. That kind of gap appears when a feature is technically possible but too expensive to deliver at the price the subscription implies. Salesforce is the most-watched case, not a unique one.

Nvidia VP of Applied Deep Learning Bryan Catanzaro put it plainly: "For my team, the cost of compute is far beyond the costs of the employees."

Should you worry about your own AI bill?

The fixes exist. Vendors can route cheap queries to smaller models and cut inference bills by around 60%. They can cache repeated instructions so the model doesn't re-read them from scratch on every call. Context discipline, trimming what's passed to the model at each step, also helps. None of these are exotic, but all require treating AI cost as a core engineering priority rather than an afterthought.

For users of business software, the practical signal is this: if an AI feature seems less capable than the demo promised, pricing pressure is a plausible explanation. Ask your vendor how they meter agent usage and whether heavy use triggers extra charges. The honest ones'll have an answer ready.

What to watch here isn't whether models get cheaper. They will. The real question is whether the per-seat SaaS model can survive contact with agentic workloads at all, and so far the evidence from vendors reporting negative gross margins suggests it cannot without a fundamental redesign.