DeepSeek Slashed Prices 75%, But AI Agent Costs Keep Climbing Anyway

The Discount That Didn't Help

DeepSeek recently cut prices on its V4-Pro model by 75%, a move that should have been straightforward good news for anyone building AI products, according to VentureBeat. Instead, many enterprise AI vendors are discovering their margins aren't improving. Some are getting worse.

Inference costs are falling, but the number of billed operations per user request is rising faster, VentureBeat reported.

Why Cheaper Tokens Aren't Saving Money

For roughly two decades, software economics followed a predictable pattern. Infrastructure got cheaper every year while applications got more capable, and margins improved as a result. Many assumed AI would follow the same script: as frontier models improved and per-token prices dropped, inference would eventually become a rounding error on the income statement, per VentureBeat.

That assumption is falling apart. A basic chatbot turns one user question into one model call, with an input-to-billed ratio of roughly 1:5, according to VentureBeat's breakdown. An AI "agent" turns that same question into a chain of steps: planning, retrieval, tool use, verification, summarization, and follow-up decisions. The user still sees a single answer. The vendor pays for the entire loop.

VentureBeat calls this the "100x problem." In a multi-step agent deployed across customer support, sales, finance, legal review, and engineering, the input-to-billed ratio can land at 1:700 or higher. Every iteration of the loop carries forward the full conversation history, tool outputs, and reasoning traces. Nothing gets dropped, so nothing gets cheaper just because the per-token price did.

VentureBeat's own walkthrough of a "simple" agent query, something as basic as asking what a top customer inquired about last week, shows the mechanics. That single sentence triggers seven priced operations: a 50-token user prompt, a roughly 3,000-token system prompt and tool-definition block repeated on every call, a 5,000-token retrieval step, three separate model calls for tool selection, summarization, and follow-up decisions, and a tool-execution return of about 4,000 tokens. By the time the answer comes back, roughly 35,000 input tokens have been billed, costing somewhere between $0.10 and $0.40, according to VentureBeat's figures.

What the Money Trail Shows

The clearest evidence of how seriously the industry takes this is what model providers are now offering developers. VentureBeat reported that OpenAI has proposed giving every Y Combinator startup $2 million in API credits, a figure that would have covered an entire seed round in past tech cycles, when the same startup cohort got by on a few thousand dollars of AWS credits.

This is an admission of what it now actually costs to run an AI-native company through its first year, per VentureBeat's reporting. For large enterprises retrofitting agents into existing products rather than building from scratch, the absolute dollar figures are bigger still.

The Fair Case for Optimism

The optimistic argument, and it's a reasonable one, is that this is simply what early-stage technology cost curves look like. DeepSeek's 75% price cut is real and it is a genuine reduction in the baseline cost of a single model call. Falling prices compound over time, and vendors building more efficient agent architectures, ones that trim redundant context and cache repeated system prompts, could bring the multiplier down without waiting on model providers to cut prices further.

That argument has merit. Prices for frontier models have fallen sharply and repeatedly across the industry. But VentureBeat's numbers show that falling per-token prices are being outpaced by rising per-task token volume, at least for now. A 75% price cut doesn't help much if the same task now requires 100 times more billable operations than it did as a simple chatbot exchange.

What's Left Unresolved

No enterprise vendor named in VentureBeat's reporting has disclosed exact margin figures for their agent products, so the scale of the profitability squeeze industry-wide remains unquantified. The open question is whether architectural fixes, trimming context, caching prompts, reducing redundant tool calls, can close the gap faster than providers can cut prices, or whether the $2 million-credit strategy becomes the new normal cost of entry for any company trying to build on top of agentic AI.