DeepSeek's V4 Flash Is So Cheap It Broke the Economics of Agent Workflows

As of August 7, 2026, DeepSeek's official API pricing page states that it will soon raise API service prices overall, with a significant increase expected, but the specific plan is still subject to formal notice. In other words, the price hike is basically confirmed, but exactly how much and when it will take effect cannot be directly concluded yet. (Models & Pricing)

My feeling is: The surge in V4 Flash usage is likely a major trigger for the price hike, but this shouldn't be simplified as "the model is too strong, so servers can't keep up." More accurately, low prices, strong capabilities, and Agent workflows appeared simultaneously, suddenly unleashing a large amount of latent demand.

V4 Flash Indeed Hits the Agent Demand Sweet Spot

When DeepSeek released V4, its positioning for Flash was clear: its reasoning ability is close to V4 Pro, achieving near-Pro performance on simple Agent tasks, while having a smaller parameter scale, faster speed, and lower price. The official documentation also specifically provided integration instructions for coding tools like OpenCode and Claude Code. (DeepSeek API Docs)

This means V4 Flash is not facing ordinary chat users, but a group of productivity users who will continuously call the model.

An ordinary user asking a single question might only consume a few thousand tokens. A Coding Agent is completely different. It needs to read project files, analyze code, call tools, run commands, check errors, and continue modifying. A single task can go through dozens of interaction rounds, and the context keeps growing longer.

If the model's capability is stable enough and the price is so low that there's almost no need to budget carefully, user behavior changes.

Previously, a developer would consider: "Should I give this piece of code to the model?" Now, it might become: "Let the model run through it first."

This is the demand amplification effect brought by low prices.

What Does the OpenCode Data Tell Us?

Public data from OpenCode can verify an important trend: DeepSeek V4 Flash has become the most-used model in OpenCode Go.

As of August 6, OpenCode data shows that V4 Flash ranks first within its statistical scope, accounting for about 59% of observed token volume, with cumulative tokens exceeding 102T and cumulative completed sessions exceeding 10.88 million. The average session consumes about 9.3 million tokens, the input cache hit rate is about 96%, and the average cost per session is about $0.09.

This data is very striking.

It shows that V4 Flash is not a model people "try occasionally," but one that is continuously called in a large number of Agent tasks. More importantly, the average session volume of 9.3 million tokens far exceeds ordinary chat scenarios.

Price Hikes May Be a Regulatory Tool, But Not the Only Way

When a model suddenly takes on a large number of Agent tasks, the platform must deal with not just total token count, but also concurrent requests, long contexts, tool calls, and sustained output.

DeepSeek's current pricing page shows that V4 Flash's cache-hit input price is 0.02 yuan per million tokens, cache-miss input is 1 yuan, and output is 2 yuan; V4 Pro's corresponding prices are 0.025 yuan, 3 yuan, and 6 yuan. Both models support 1 million token contexts, but Flash has a higher concurrency limit.

For developers, such pricing encourages heavy use of caching and long contexts. For the platform, caching can reduce some computational costs but cannot eliminate all pressure. Long-chain Agents will still continuously initiate requests, persistently occupying inference resources.

Therefore, a price hike can indeed serve two purposes:

First, reduce invalid and excessive calls. Second, prioritize limited computing power for users willing to pay higher prices.

What Should Users Really Care About?

For individual users who call the API occasionally, the actual impact of the price increase may not be as large as imagined. The truly sensitive users are those using Coding Agents, automation scripts, batch processing, and long-text analysis.

These users cannot just look at "how much per million tokens," but also need to consider:

How many interaction rounds the model needs to complete a task; How many tokens are consumed per task on average; Whether the cache hit rate is stable; Whether the service is reliable during peak hours.

If a model's price is very low but it frequently times out, fails, or repeats calls, the final cost may not be low. Conversely, a model with a slightly higher price but that can complete a task in one go may have a better overall cost.

So, after DeepSeek's price hike, V4 Flash may still be competitive. As long as its success rate in code modification, tool calls, and long-context tasks remains leading, it won't necessarily lose users immediately because of the price increase. What users will really compare is "how much does it cost to complete a task," not simply the input token price.

What Will This Price Hike Change?

I think what this event really changes is the market's understanding of large model pricing.

In the past, many people viewed low prices as a value that model manufacturers must adhere to long-term. In reality, low prices often serve the purpose of acquiring users, expanding the ecosystem, and testing load. When models enter the Agent phase, and users shift from "Q&A" to "delegating work," call volume rapidly amplifies, and the original low prices may become difficult to sustain.

What DeepSeek faces is not a simple price hike choice, but a commercialization stress test.

If it explains the capacity issues in advance, announces a clear pricing gradient, and leaves a transition period for existing users, users may accept this adjustment. If it just suddenly raises prices significantly without explaining service quality, concurrency capacity, and subsequent guarantees, user trust will indeed decline.