AI Agent Model Pricing! Why DeepSeek API Prices Surge While OpenAI Cuts 80%?

A new shift is emerging in the AI model market's pricing wars. While competition among large models has centered around 'capability improvement and price reduction' over the past two years, recent pricing strategies among companies are beginning to diverge: DeepSeek has raised API prices, and Kimi K3 has entered the market at higher prices for long-term programming and complex knowledge work. Meanwhile, OpenAI has significantly reduced prices for some models, even making some models handle low-cost, high-throughput execution work.

Analysis shows that whether large models are 'becoming cheaper or more expensive' can no longer be judged solely by the price per million tokens.

As AI agents begin to handle longer, more complex tasks, what truly affects enterprise costs is gradually shifting from model prices to how many model calls, how much context, and how much inference and retry costs are needed for a single task.

DeepSeek has formally adjusted API prices, with V4 Pro's non-cached input price during peak hours rising from 3 yuan to 9 yuan, and output price from 6 yuan to 27 yuan. V4 Flash has also been adjusted with peak and off-peak pricing introduced for the first time.

The most notable adjustment is that V4 Pro's output price during peak hours has reached 4.5 times the original price, and the cached input price has reached 12 times the original. DeepSeek, which previously established market recognition through its 'price slasher' image, now sees its price advantage less pronounced after this adjustment.

However, even after the adjustment, DeepSeek's prices still have a gap compared to some top U.S. models. This makes this price adjustment more like a test of existing usage scale and customer price sensitivity.

Another Chinese model company, Moonlit Face, recently launched Kimi K3, adopting a different strategy. K3 is mainly aimed at long-term programming, agents, and complex knowledge work, with a standard input price of 20 yuan per million tokens and an output price of 100 yuan.

This price is significantly higher than most Chinese open-source models, but it is not a sudden price increase for an existing low-priced product.

From DeepSeek raising the prices of existing services to Kimi K3 entering the market at higher prices, the market image of Chinese model companies, which have relied on low prices, is beginning to change.

Meanwhile, OpenAI has taken the opposite approach.

On July 30, OpenAI reduced the API price of GPT-5.6 Luna by 80%, and Terra by 20%. The quotas consumed by ChatGPT Work and Codex using these two models were also reduced in sync. The standard price of Sol was maintained, but the Fast mode for low-latency needs was changed to charge at twice the standard price.

This seemingly contradictory pricing strategy actually reflects the differentiation of model roles in AI workflows.

OpenAI provided an example of a programming task in model division: Sol is responsible for understanding the problem, handling uncertainty, and formulating plans, while Luna is responsible for executing already determined modifications, writing tests, executing tests, and checking results.

In other words, Sol is responsible for deciding 'how to do it,' while Luna is responsible for executing 'what has already been decided to do.'

Their usage frequency and value in a single task are not the same. High-end models may only need a few calls to determine the direction of the entire task, while low-end models may continuously execute dozens of times in the background. Therefore, model costs will directly affect whether agents can operate on a large scale.

In this architecture, Luna is not just a cheaper model, but may also become the execution layer for OpenAI to handle large-scale agent workloads.

After the price reduction, tasks that were previously too costly to execute may become economically viable, and agents already in operation can increase testing, checking, and retry times.

Therefore, OpenAI has reason to lower Luna's price. As long as more workloads stay within Codex and OpenAI API, some high-value steps may still be upgraded to Terra or Sol. Low-cost models handle a large number of tasks, while high-end models retain higher prices in key steps.

The bigger change behind this is that AI agents are changing the way models are used.

Traditional chat products are typically question-and-answer based, but agents, after receiving a task, may need to first read documents and historical records, then formulate plans, call tools, perform operations, and check results. If errors occur during the process, they may need to re-read the context, modify the plan, and execute again.

Therefore, a single user instruction may be transformed into dozens or even hundreds of model calls. This makes the price per million tokens no longer sufficient to measure the actual cost of an AI task.

Enterprises need to focus on how many tokens, how many model calls, how much context, and how many retries and human interventions are needed to complete a task.

Microsoft's research team evaluated 8 reasoning models in 2026 and found that in 21.8% of pairwise model comparisons, the lower-priced model had a higher total cost, with a maximum difference of 28 times. The main reason is that different models may consume vastly different amounts of thinking tokens to solve the same problem.

McKinsey's internal usage experience also shows that the cost of the same agent programming task may vary by 30 times, and about 10% of users consumed 65% of the tokens. Therefore, the institution began using model routing, leaving more expensive models for difficult tasks and daily tasks to older or smaller models.

This means that model price reductions do not necessarily mean a corresponding reduction in enterprise AI spending.

If the unit price of tokens continues to decline, enterprises will have more budget space to use AI, and lower costs will encourage more tasks to be handed over to agents for processing.

Thus, a seemingly contradictory phenomenon emerges: models are becoming cheaper, but the 'intelligence' consumed by each user may continue to increase.

OECD statistics show that from January 2024 to April 2026, the price of text models, after quality adjustment, fell by nearly 80%. Now, some small models can provide similar capabilities at a cost far lower than that of flagship models a year ago.

However, on the other hand, the tasks that agents undertake are becoming longer, with more execution steps, and additional checks, retries, and tool calls are needed.

Epoch AI estimates that global reasoning supply capacity increases by more than 3 times annually, but some demand indicators may approach an annual increase of 10 times. Long-context agents may face capacity limitations first.

This may also lead the market to form a more distinct model division of labor: ordinary work continues to shift to smaller, cheaper models, while users willing to pay for the most cutting-edge capabilities may need to bear higher prices.

OpenAI's current pricing structure is showing this differentiation.

Luna and Terra are responsible for low-cost, high-throughput work, while Sol retains a higher capability premium. The Fast mode, which charges twice the standard price, reflects the scarcity of 'instant computing power' as a resource.

When users require models to complete tasks immediately, service providers need to provide higher instant processing capabilities and cannot simply reduce costs through queuing and batch processing.

Therefore, the model market is beginning to show a pricing structure similar to cloud computing or even the power market: basic loads are becoming cheaper, but peak capacity, priority processing, and low-latency services can still command a premium.

From this perspective, the significance of DeepSeek's recent price adjustment is not just to simply increase API revenue.

DeepSeek has quickly expanded model usage scale through low prices, allowing developers and enterprises to build an ecosystem of model access, fine-tuning, and tool adaptation.

Once users have embedded models into existing workflows, whether price increases will cause a large number of customers to switch to other models can reflect whether DeepSeek relies on low prices to attract users or has established sufficient product stickiness.

Previously leaked investor meeting minutes mentioned that DeepSeek's API pricing target is to recover device costs in about 10 months. Under this premise, if the original price is already below the price sensitivity range of most users, further price reductions may not bring proportional new demand, but may instead compress revenue.

Compute resource limitations are also an unavoidable factor.

The same exchange materials pointed out that DeepSeek's access to domestic chip resources is significantly less than that of large internet companies. Although model training and inference do not necessarily always use the same equipment, raising API prices may, through screening some low-value or delayable loads, improve cash flow and allow limited resources to be used more effectively.