Over the past year,the large model industry has been embroiled in a brutal price war,a trend that shows no signs of dissipating as the new year dawns.The logic behind this fierce competition is relatively straightforward: lowering prices serves as both a market education tool and a survival strategy.However,the ultimate victor in this battle will be determined by technological prowess,ecosystem strength,and user preferences.

On the final day of 2024,Alibaba Cloud made headlines by announcing the third round of price reductions on large models,showcasing a staggering drop of over 80% on its Tongyi Qianwen visual understanding model.This isn't an isolated incident; just weeks prior,Volcano Engine revealed at the Force conference that their Doubao visual understanding model had its input price slashed to 0.003 yuan per thousand tokens,translating to merely 1 yuan for processing 284 images at 720P resolution.Such unbelievably low prices cast doubts on the existence of profit margins.In an even bolder move,Baidu's intelligent cloud proclaimed that two core models from its Wenxin series would be entirely free for users.

However,price cuts come at a cost.These actions further squeeze the survival space for smaller players.For tech giants,lowering prices is a tactic to build ecological barriers,while for smaller model providers,dwindling profits and insufficient R&D funding may lead to stagnant technological advancement,ultimately pushing them out of the race.The intensifying competition indicates that this price war is actually a struggle for survival in the industry.

The price reductions in large models represent a complex game of commercial maneuvering hidden behind the surface of apparent rivalry.To truly understand the strategies at play,one must dissect their underlying business models.At its core,the battle revolves around basic services,refined needs,and exclusive computing power,with price reductions merely being the facade of a deeper,intricate game.

Current market research suggests that large model vendors can be categorized into three primary business model types:

1.**Basic Services**: This aspect pertains to the most direct user interactions,where input content generates output during the model inference process.Consequently,it bears the lowest pricing thresholds,with price wars predominantly centered here.

2.**Model Fine-Tuning**: Vendors tailor models according to customer requirements,charging based on the token count of training texts and the number of training iterations.These charges are relatively steep and typically follow a pay-as-you-go model.

3.**Model Deployment**: For large clients,vendors allocate dedicated computing resources,charging based on actual computing consumption or token usage.This offering is more akin to personalized customization with a complex pricing logic and substantial premiums.

These three models illustrate the service pathway extending from basic to advanced levels.The current focus of price reductions primarily revolves around the first service category: the cost of basic model inference.

However,these price cuts aren’t arbitrary; rather,they are meticulous strategies.Take Doubao's general model Pro-32k,for instance: its input cost is set at 0.8 yuan per million tokens,a staggering 99.3% lower than the industry average.Yet,the output cost soars to 2 yuan per million tokens—nothing particularly remarkable in comparison to competitors.This “low input,slightly higher output” pricing model is designed to entice developers into trial use while ensuring profit margins remain intact.

Turning to the visual understanding model,Doubao-Vision-Pro-32k has its input priced at 3 yuan per million tokens,with output reaching as high as 9 yuan per million tokens.Though it claims an overall price reduction of 85%,it doesn’t offer a clear advantage over direct competitors like Alibaba's Qwen-VL or Gemini 1.5 Flash.

Other domestic vendors exhibit similar tactics.Baidu's ERNIE-Speed-8K,while marketed as "free," imposes a charge of 5 yuan per million tokens upon actual deployment.Similarly,Alibaba’s Qwen-Max and Doubao’s Pro-32k significantly compete only on basic services while maintaining high prices for advanced offerings.

On the surface,these price cuts appear to be aimed at capturing users; however,they are fundamentally strategic moves by vendors to secure market influence.By offering drastically low prices,they lower the trial barrier for small and medium developers,leading to an accumulation of user data and optimization of model performance—thereby setting the stage for high-end services in the future.

Nevertheless,the real profit-driving segment does not lie within these lightweight basic services but instead is concentrated on custom training and exclusive computational deployments.For instance,Doubao's Pro series commands a fine-tuning service price as high as 50 yuan per million tokens,exponentially higher than basic services—far exceeding the market's price reductions.

Though major players are prepared to sacrifice profits to seize market share,this approach does come with its own risks.The competition for low-priced lightweight models may erode overall industry profit margins,potentially leaving smaller vendors out of the race.Ultimately,users are most concerned with model performance and practical application efficacy.

A key question arises: will the surge in user numbers from price cuts translate into sustained revenue gains?If users remain fixated solely on basic services without exploring fine-tuning or deployment options,the expenses associated with pricing reductions may never be recuperated.

The ongoing price war in large models resembles a free-to-play online game: easy to join,but for more powerful gear and skills,one must continuously invest money.Vendors lure users with low prices while essentially building a reservoir for high-end services.However,the long-term effectiveness of this strategy hinges on whether users are willing to pay for performance enhancements,alongside the vendors' capacity to maintain technical competitiveness amidst price cuts.

Another prominent player in this space,ByteDance,grapples with its own anxieties regarding what it perceives as an "interim state." Upon reviewing the large model landscape,it's apparent that ByteDance was not among the early frontrunners.CEO Liang Rubo acknowledged internally that ByteDance's responsiveness to large models was "delayed," especially when compared to startup firms established between 2018 and 2021—in fact,significant discussions around GPT only commenced in 2023.

The latecomers often need to compensate for their initial lag through intense competition,and ByteDance’s strategy has been to "invest heavily to generate buzz." In the consumer market,Doubao has swept in with an all-encompassing presence both online and offline.Reports indicate that by the first half of 2024,ByteDance aggressively ramped up capital injections into advertising for Doubao,pouring between 15 to 17.5 million yuan during April to May,with expenditures soaring to 124 million yuan by early June.Furthermore,ByteDance capitalizes on the robust traffic of its Douyin platform,effectively sidelining other AI applications,thus positioning Doubao as a standout product in resource acquisition.

According to AppGrowing,by November 2024,Doubao and another native AI application,Kimi,emerged as the two most aggressive products in terms of advertising spending in China,investing 540 million yuan and 400 million yuan respectively.This costly marketing approach propelled Doubao to become the AI software with the largest user base in the country,yet the associated concerns are glaringly evident.

Despite hefty advertising investments,user retention has not met expectations.Statistics reveal that Doubao’s user activity is relatively low,with engagement lasting merely 2 to 3 days per week,users sending only 5 to 6 messages daily and averaging a meager 10 minutes of usage time.These metrics have not shown significant growth over the past year,indicating a low level of actual user reliance on the product.

Management at ByteDance exhibits a cautious internal assessment,regarding Doubao as potentially just a "transitional state" among AI conversational products.In China,subscription-based revenue models struggle to gain traction,and with low user engagement and interaction frequency,the potential for advertising revenue development faces significant constraints.This dual dilemma places an invisible ceiling on Doubao's commercialization trajectory.

The bottleneck in the consumer market has compelled ByteDance to devote more resources toward B2B operations.From slashing prices to vigorously promoting Doubao's multimodal models,ByteDance seeks to establish a solid cash flow through enterprise-level services.However,navigating the Chinese market significantly diverges from the US landscape,making it challenging for AI products to achieve profitability solely through subscription or platform sales.More often,the success of commercialization hinges upon specific project deployments,which are frequently tied to the vendor's stature and the trust they command in the market.

This reality highlights the necessity for newcomers and established giants alike to thrive amidst industry enthusiasm.Startups seek to create buzz for fundraising purposes,while corporations like ByteDance strive to leverage that excitement for customer acquisition.Nonetheless,such strategies carry inherent risks: enthusiasm does not directly translate into sustainable revenues,especially when the product's intrinsic user value remains underdeveloped.

From ByteDance's shifting strategy perspective,an emphasis on “lower barriers,more modalities” could represent a potential breakthrough.Video models like Jianying and Jiemeng may emerge as critical scenarios for successful large model implementations.Doubao's recent focus on the video sector at the Force conference underscores the potential the company sees in this field.

However,it's crucial to realize that the ultimate decision-maker remains the user.

ByteDance’s path choices are indicative of thoughtfulness but simultaneously expose a common industry issue: an excess dependency on trends while overlooking the genuine accumulation of user value.The current status of Doubao reflects a crucial truth: merely "pouring money" into attracting users cannot compensate for deficiencies in product experience.Furthermore,the anxiety regarding being trapped in a transitional state isn’t unique to ByteDance; it encapsulates the entire large model industry’s struggle to find a balance amidst the ongoing price war and user experience improvements.

As the price war intensifies to boiling point,the race for superior technology commences.The large model sector has never been a gentle battleground.In 2024,as the market reshapes itself,the elimination competition dubbed "survive the last one" has pushed many second-tier vendors to the sidelines.Survival in this landscape is now more than just a matter of price—a matter of technology prowess.

Beginning in 2023,lower-tier models have resorted to cash-burning methods to retain users.This blunt approach—cutting prices,offering free services,and amplifying ad spending—merely aims at capturing C-end users and seizing market share,yet proves increasingly unsustainable.

The customer acquisition costs within the domestic large model application sphere are soaring rapidly.Even the most aggressive spenders find that free product offerings do not convert into actual profits.Concurrently,user engagement at the consumer level remains lackluster,with high attrition rates.Crucially,enterprises must target B-end clients willing to pay—sectors such as finance and government,where high added value is commonplace—and these clients prioritize technology and service capacity over pricing.

As the market stabilizes,the simplistic strategies of a price war gradually begin to cede ground to a race focused on technology advancement.Who can deliver superior technological capabilities at lower costs will emerge as the ultimate survivor in this competition.

The reduction of technological costs hinges on hardware capabilities and algorithmic optimization,with domestic mainstream vendors currently exhibiting minor discrepancies in this domain.Consequently,a differentiated technological path may become key to breakthroughs.Recently,the much-discussed DeepSeek-V3 showcases how distillation technology can streamline and optimize generative AI performance,offering new explorations for the industry.However,this progression has sparked debates concerning the merits of “optimizing GPT,” with experts divided on its true value.

Every significant shift in technology begins moving the focal point from “generative AI” to “inferential AI.” OpenAI's launch of the o1 model serves as an industry lighthouse; this product represents a cornerstone leap from generative to inferential capabilities in AI by mimicking human logical reasoning through delayed inference techniques.Even more striking,just three months after releasing o1,the iterative o3 model swiftly emerged,enhancing reasoning speed and task adaptability,alongside introducing a mini version—highlighting OpenAI’s dominant pace in technological evolution.

In contrast,while domestic vendors have made strides in catching up,the gap between them and OpenAI remains stark.Efforts to enhance reasoning capabilities through integrated chain thinking,tree search mechanisms,and reflective strategy optimization have yet to surpass the advancements seen in the o series.According to an industry expert,securing access to open-source models supporting the o1 architecture may propel domestic vendors forward,yet existing technology faces numerous bottlenecks before this becomes a reality.

Historical patterns within the GPT series indicate a gap of roughly one year from a leading vendor's revolutionary tech launch to broader industry adoption.By the first half of 2025,it is likely that domestic vendors will complete their preliminary catch-up to the o series.Post that,generative AI technologies may gradually fade from prominence.

For domestic large model vendors,time is of the essence.In the ensuing elimination round,the efficacy of pricing strategies will diminish,leaving technological capabilities as the solitary bargaining chip for survival.Instead of waiting for obsolescence,it is imperative to capitalize on the existing momentum of generative AI,enhancing product performance and optimizing service experience to secure a competitive edge.