Leveraging the Pareto curve, are large models redefining cost-performance ratio?

DeepSeek has officially launched the new version V4.1 Flash-1017 of its large model. Public data previously released has shown its technical indicators, and this time DeepSeek has for the first time independently released the Pareto frontier curve, which intuitively presents the model's comprehensive performance in the dimensions of efficiency and cost.

V4.1 Flash-1017 has reached the industry-leading level in this curve: the comprehensive cost per unit task is only 1/100 to 1/20 of that of similar models such as Fable and Astra, with a cost-performance advantage of more than 40 to 50 times.

Actual measurements show that in most scenarios, Token consumption has decreased by more than 50% compared with the previous version, and the corresponding TCO level will also usher in a new round of substantial decline simultaneously.

An update whose suffix looks like a minor version iteration may be rewriting the way the entire industry calculates accounts.

The era of bargain prices is over, and the bill is starting to speak

Over the past two years, domestic large models have experienced an extremely fierce API price war.

In May 2024, Doubao large model focused on "extreme cost performance" in terms of pricing, with the inference input price of the general model pro-32k version at 0.0008 yuan per thousand tokens, 99.3% lower than the industry average price.

Subsequently, Alibaba Cloud issued a price reduction announcement, sharply cutting the prices of a number of commercial and open-source models of its Tongyi Qianwen. Among them, the API (Application Programming Interface) input price of Qwen-Long, a commercial model benchmarking GPT-4, was reduced from 0.02 yuan per thousand tokens to 0.0005 yuan per thousand tokens, a direct drop of 97%; the output price was reduced from 0.02 yuan per thousand tokens to 0.002 yuan per thousand tokens, a direct drop of 90%.

Baidu Intelligent Cloud also released an announcement stating that two flagship models of Baidu ERNIE large model, ENIRESpeed and ENIRELite, will be fully free of charge with immediate effect.

Domestic large model APIs once entered the "cent era", and for some lightweight models, users can buy millions of tokens with 1 yuan on the input side.

However, there are rigid constraints on underlying computing power and inference costs, and prices cannot converge infinitely towards zero. Since mid-March, China's cloud and AI computing power market has ushered in a round of structural price shocks: the prices of computing power infrastructure and some model APIs have been raised, and the industry's pricing logic has begun to diverge.

On March 18, Alibaba Cloud announced that due to the explosion of global AI demand and rising supply chain prices, the procurement cost of core industry hardware has risen significantly. The company decided to adjust the prices of AI computing power, CPFS (Intelligent Computing Edition) and other services from April 18, 2026. Among them, the computing power cards of the XuanTie series of C-Sky increased by 5% to 34%, and CPFS (Intelligent Computing Edition), a file storage product, increased by 30%.

Tencent Cloud raised the input price of the Tencent HY2.0 Instruct model by as high as 463%. It also announced that it will uniformly raise the prices of AI computing power, container services and elastic MapReduce related products by 5% from May 9. Baidu Intelligent Cloud announced that from April 18, the prices of AI computing power related products and services will be increased by about 5% to 30%.

In the overseas market, Amazon AWS has implemented a 15% price increase for EC2 instances used for large model training, and the price of the top-tier instances equipped with H200 chips has increased by 15%; Google Cloud even announced that the price of AI computing instances will be increased by 20% to 50%, and the price of CDN egress bandwidth in some regions has increased by as high as 100%.

Behind the price increase is a clear business logic: it is difficult to achieve positive gross profit by relying on continuous subsidies; at the same time, enterprise customers' positioning of large models has changed. They are no longer just tools for testing and experience, but gradually embedded in business links as production resources.

What really leverages the change of the industry's competition logic is the iteration of the task form carried by large models.

When the model shifts from answering questions to executing complete projects in a closed loop, the cost structure changes accordingly: the programming agent needs to go through the whole process of requirement analysis, tool call, code rewriting, test verification and loop repair.

Industry measured data shows that unoptimized native prompts have 30%-55% invalid Token consumption, and long-term high-frequency calls will cause a large amount of waste of computing power resources and funds.

After the downward space of unit price reaches the ceiling, how to spend money is more critical than simply looking at the unit price of Token. The foundation for the shift of the main axis of competition is gradually taking shape, and a Pareto curve is also beginning to reconstruct the industry's evaluation criteria for cost performance.

The Pareto frontier is not a new concept. In layman's terms, it is the optimal boundary between performance and cost: every point on the boundary follows the trade-off logic — if you want to further improve performance, you have to bear higher costs; if you want to reduce expenses, you have to accept performance loss.

Solutions that fall inside the boundary still have room for Pareto improvement: there is an opportunity to reduce costs without sacrificing performance, or to increase performance without reducing costs. The essence of this DeepSeek upgrade is a typical Pareto improvement. The new version moves closer to the frontier curve, achieving both performance improvement and lower unit task cost.

In the past, most manufacturers' common practice was to display indicators separately: releasing benchmark scores on one side and listing API price lists separately on the other, rarely putting the two in the same coordinate system for comprehensive trade-off.

The special feature of this DeepSeek move is that as a model provider, it actively released the performance-cost Pareto frontier curve for commercial scenarios, which intuitively presents the comprehensive trade-off relationship of the model in the two dimensions of efficiency and cost.

In the DeepSWE v1.1 evaluation chart released by DeepSeek, the highest pass rate (Pass@1) of the new version reaches 77.0%. In horizontal comparison, the highest point of the previous version DeepSeek V4.1 Flash is about 74%. The highest score of the new version appears at the position where the average cost per task is about 1.3 yuan, which is significantly lower than many overseas models with similar performance in the chart.

The business implication of this incident is far more than that model calls are cheaper. Establishing the cost per unit task as a scale for horizontal comparison essentially puts forward a new evaluation framework: in the past, B-side selection was more likely to prioritize the scores of public benchmark lists, and this new framework will push purchasers to compare the actual expenses behind the completion of the same business goal.

However, this curve is drawn by the manufacturer itself: the supporting task set, evaluation framework DeepSeek Harness, and inference gear rules are all internally defined. Whether the cross-model comparison is valid and whether the outside third party can reproduce the cost gap of 1/100 to 1/20 remains to be further verified.

At least for now, it is closer to an official statement given by the manufacturer, rather than a conclusion that has undergone independent third-party audit.

The marked price remains unchanged, and the bill shrinks: a flexible cost reduction

The most intriguing design of the DeepSeek V4.1 Flash-1017 iteration is that the external API call method and the public pricing system remain completely unchanged, with no price reduction.

The cost dividend brought by the model is not released by directly modifying the price list, but hidden in the underlying reconstruction of Token usage efficiency.

In the interpretation for commercial customers, this iteration summarizes Token optimization into three layers of mechanisms: task-aware compression, dynamic context clipping, and adaptive decoding strategy; the underlying layer corresponds to a series of technical combinations such as KV cache compression, context management, and decoding scheduling.

Actual measurements show that in most scenarios, Token consumption has decreased by more than 50% compared with the previous version, and the corresponding TCO level will also usher in a new round of substantial decline simultaneously.

This model can be regarded as a flexible price reduction idea in large model services: the external public marked price remains unchanged, but the actual call bill corresponding to customers completing the same task can shrink. Compared with directly reducing the API unit price, this strategy skillfully avoids multiple risks associated with direct price wars.

First of all, it is not easy to directly trigger a new round of public price wars. Directly reducing the API unit price will generate a transparent and easy-to-follow digital anchor point, and competitors can quickly follow suit to reduce prices. Eventually, the industry will easily fall into unit price involution, collectively squeezing the gross profit margin.

The cost reduction brought by efficiency optimization is deposited in the task execution link, which belongs to the hidden advantage at the level of model engineering, and the threshold for competitors to replicate and verify benchmarking will be higher.

Secondly, the existing contract pricing system of existing customers will not be impacted. For enterprise customers who have already signed contracts, the public API unit price remains the same, and the original business terms and budget calculation framework do not need to be readjusted; customers can obtain lower actual task costs in a nearly imperceptible state, which helps to stabilize customer retention.

Going a step further, the manufacturer's competition narrative is also changing. Competitive advantage is no longer limited to the intuitive label of "how much per thousand Tokens", but becomes a composite indicator formed by task completion rate and unit task cost. The focus of industry price comparison is gradually shifting from the unit price of Token to how much actual expense is ultimately required to achieve the same business goal.

High-frequency call scenarios, Agent workflows, batch automated tasks, and businesses like RAG Retrieval-Augmented Generation are often high-incidence areas of redundant Token consumption. The repeated back-transmission of context and the redundant and lengthy output of models continue to push up the comprehensive cost of long-term calls.

The significant drop in Token consumption this time can optimize the cost structure of such businesses, broaden the landing boundary of Agent workflows, and make many automated businesses have landing possibilities in terms of budget.

Looking back at the first launch of V4.1 Flash, DeepSeek chose to directly lower API pricing and launch a peak-valley billing scheme at that time, with half price for off-peak calls. More than a month later, the 1017 version no longer adjusts the public marked price, but releases dividends relying on engineering optimization of the inference layer and Token link.

This change conveys a noteworthy signal: the marginal benefit of the direct price reduction path is declining, and the focus of industry competition is gradually shifting from the digital game on the price list to underlying engineering capabilities such as inference efficiency and Token link optimization.

The money saved on Tokens is not free

The cost dividend released by Token efficiency optimization has clear applicable boundaries.

Scenarios with relatively significant benefits are concentrated in businesses with high-frequency calls, relatively short context, and structured output, such as programming agents, Agent workflows, RAG Retrieval-Augmented Generation, and online customer service.

The overhead bottleneck of such scenarios often does not lie in the basic inference cost of the model itself, but in the invalid Token consumption caused by redundant output and repeated back-transmission of historical context in multi-turn conversations. Technical means such as task-aware compression and dynamic context clipping are precisely aimed at this pain point, and the leverage effect of cost reduction will be more prominent.

However, the industry needs to recognize a reality: the cost savings brought by Token compression are never cost-free.

When model output streamlining and inference link compression are applied to complex inference, ultra-long document parsing, and multi-step chained logic tasks, potential risks will emerge: is it possible for the compression strategy to lose key details and cause inconsistent inference results?

The average task pass rate obtained by relying on official benchmark evaluation is difficult to cover such hidden dangers. Being able to achieve streamlined output in the test set does not mean that it can still maintain key information and result reliability when deployed to diverse real businesses. Compression is a trade-off between information retention and inference cost. Once the balance point shifts, behind the cost reduction may be the loss of business quality.

At the same time, a group of unavoidable controversies falls on the comparison caliber. The official conclusion that the comprehensive cost per unit task is only 1/100 to 1/20 of similar models such as Fable and Astra is based on prerequisites: Flash focuses on the lightweight high-speed route, while Fable 5 and Astra are flagship models, whose positioning, computing power base and upper limit of capabilities are not on the same baseline.

The fairness of horizontal cross-tier model cost comparison needs to be carefully disassembled, and the multiple difference cannot be directly equated to an all-round cost performance crushing.

When OpenAI explained Astra to the public, it also emphasized optimizing cost-effectiveness with fewer Tokens and lower retry frequency. Manufacturers at home and abroad are sharing the same narrative of optimal task cost, but the evaluation systems and task datasets of all parties are not interoperable, making horizontal benchmarking very difficult.

This also leads to a noteworthy hidden worry in the industry: if the Pareto frontier curve gradually becomes a standard tool for manufacturers' external marketing, will this performance and cost measurement framework give rise to a new round of list involution?

Task datasets, inference gears, and evaluation operation frameworks all have adjustable space, and every manufacturer can draw an impressive-looking frontier curve in a self-defined coordinate system. More and more self-contained evaluation standards will amplify the information noise in the industry instead.

For B-side purchasers, the Pareto curve released by manufacturers is only suitable as reference material, and the most reliable scale to measure model value is still the actual test in their own business scenarios.

First, the main line of competition in the API market may shift from paper marked prices to effective Token efficiency. The core of manufacturers' competition in the next stage may not be the unit price per million Tokens displayed on the official website, but how many effective business results the model can stably deliver under the same budget. The reference value of unit price figures may continue to decline, and the input-output ratio per unit task is expected to become a new core indicator.

Second, the B-side selection logic may also be reshaped accordingly: building internal benchmarks based on own business data to measure the cost of a single task and task completion rate may gradually replace the general list and become the primary link of model evaluation. Purchasers no longer blindly believe in public scores, and pay more attention to the measured performance in business scenarios.

Third, the competition in the lightweight Flash track may further intensify. DeepSeek and MiniMax have taken unit task cost as their core selling point, and domestic lightweight high-speed models continue to increase investment, which may continuously push down the price anchor of similar overseas products. The Flash track is no longer just a simple low-cost alternative, and begins to compete for the deployment share of Agent automation scenarios head-on.

Fourth, engineering scheduling capabilities may obtain the same weight as the basic capabilities of the model. Engineering optimization methods such as context management, inference scheduling, and Token compression strategies are expected to gradually become differentiated barriers as important as model architecture and basic algorithms.

In this iteration, DeepSeek has simultaneously opened its self-developed Search and Fetch API, and completed the adaptation with Agent frameworks such as Tencent CodeBuddy and WorkBuddy, which also sends a clear signal: the competition boundary has jumped out of the model itself and extended to the full link of complete task delivery.

The prerequisite for the establishment of this business logic of reducing costs through Token optimization is that efficiency optimization cannot damage task quality. Once the compression mechanism makes frequent errors in long-tail complex scenarios, the Token costs saved in the early stage are likely to become hidden expenses brought by retries, manual review, and business failures, pushing up the total TCO of the project.

The core change brought by the discussion of the Pareto frontier curve is not the surge of a single evaluation indicator, but the transfer of the subject evaluating model value.

It can be seen that the right to judge large models is gradually breaking away from labels such as list scores and paper pricing, and shifting to the real business bills in the hands of enterprise customers.

This article is from the WeChat official account "Silicon Carbon Variable", written by Wen Ning, and authorized for release by 36Kr.