跪拜 Guibai
← All articles
DeepSeek

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

By 勇宝趣学前端 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The price hike signals that agent-driven inference loads break the unit economics that worked for chat. Developers running coding agents, automation scripts, or batch processing need to stop comparing per-token prices and start measuring cost-per-completed-task, because a cheap model that retries five times costs more than an expensive one that succeeds on the first call.

Summary

DeepSeek's V4 Flash model has become the most-used model on OpenCode, accounting for 59% of observed token volume with over 10.88 million completed sessions. The average agent session consumes 9.3 million tokens at a cost of just $0.09, a usage pattern far beyond ordinary chat that has triggered a capacity crunch. The upcoming price increase is less about raw popularity and more about the demand amplification that occurs when a model is cheap enough, capable enough, and targeted directly at autonomous coding agents. Developers who run long-chain agent tasks will feel the impact most, but the real shift is in how the market now understands model pricing: the low prices that built an ecosystem cannot survive the transition from Q&A to delegated work.

Takeaways
DeepSeek confirmed an API price increase is imminent, with the final amount and date still pending formal notice.
V4 Flash accounts for 59% of token volume on OpenCode, with 10.88 million sessions completed and an average of 9.3 million tokens per session.
The average V4 Flash agent session costs about $0.09, with a 96% input cache hit rate.
V4 Flash's current pricing is 0.02 yuan per million cached input tokens, 1 yuan for uncached input, and 2 yuan for output.
Agent workflows consume orders of magnitude more tokens than chat because they read project files, call tools, run commands, and iterate through error correction across dozens of rounds.
Low prices changed developer behavior from selective model use to running the model on everything by default.
Cache hit rates reduce compute cost but do not eliminate the load from sustained, long-context agent requests.
A price hike serves two functions: curbing excessive calls and prioritizing compute for higher-paying users.
Users should evaluate models by total task-completion cost, not per-token price, especially when reliability and first-pass success rates differ.
The event marks a commercialization stress test where the low prices that built an ecosystem become unsustainable once users shift from Q&A to delegating work to agents.
Conclusions

V4 Flash's pricing created a demand amplification loop: the model was so cheap and capable that developers stopped deciding whether to use it and started using it by default, which made the pricing unsustainable.

The 9.3-million-token average session length is the key number. It proves agent workloads are not just larger than chat; they are a different category of consumption that pricing models designed for chat cannot absorb.

Cache hit rates above 95% look like an efficiency win, but they mask the real problem: long-chain agents still generate continuous, stateful request streams that keep inference hardware occupied regardless of cache savings.

DeepSeek's pricing page still lists V4 Flash at dirt-cheap rates while simultaneously warning of a large increase, which suggests the decision is reactive rather than part of a planned pricing ladder.

The real comparison metric after the hike will be cost-per-completed-task, not cost-per-token. A model that fails and retries three times at half the token price is more expensive than one that succeeds immediately.

Agent workloads turn model pricing from a simple commodity comparison into a reliability-and-success-rate equation, which benefits providers who can demonstrate consistent first-pass completion.

Concepts & terms
Agent Workload
A usage pattern where an AI model autonomously performs multi-step tasks — reading files, calling tools, running commands, checking errors, and iterating — rather than responding to a single prompt. Each task can span dozens of interaction rounds and consume millions of tokens.
Cache Hit Rate
The percentage of input tokens that the model can reuse from a previous computation instead of processing from scratch. A high cache hit rate lowers compute cost but does not reduce the number of requests or the sustained load on inference hardware.
Demand Amplification
The effect where a sufficiently low price and high capability cause users to stop rationing a service and start using it indiscriminately, multiplying total consumption beyond what incremental growth models predict.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗