"Doubao" lowers price points as global large models compete on cost-effectiveness
TechFlow Selected TechFlow Selected
"Doubao" lowers price points as global large models compete on cost-effectiveness
Before Doubao made its debut with high cost-effectiveness, many domestic large models such as Qwen, Zhipu AI, and DeepSeek had already started competing aggressively on price.
Navigating Web3 tides with focused insightsAuthor: Mu Mu
Large models have entered a price war.
On May 15, Volcano Engine, ByteDance’s AI subsidiary, launched the Doubao large model. In addition to offering free access via the consumer-facing Doubao app, Volcano Engine slashed enterprise pricing to the lowest in the industry.
According to Tan Dai, CEO of Volcano Engine, the core Doubao model (≤32K) is priced at just ¥0.0008 per thousand tokens—only 0.8 fen—to process over 1,500 Chinese characters, representing a 99.3% cost reduction compared to competitors.
Even before Doubao's high-value debut, domestic large models such as Qwen, Zhipu AI, and DeepSeek had already begun slashing prices, pushing the "hundred-model battle" into a new phase of collective price cuts. As Tan noted, reducing costs is a key factor in accelerating large models into the "value creation stage."
Doubao Drives Enterprise Pricing to Industry Lows
The Doubao large model evolved from the earlier Yunque (Cloud Sparrow) model—the first Transformer-based large model launched by ByteDance in August 2023. Within half a year, Doubao has expanded into a full suite of offerings while significantly lowering prices for enterprise clients.
The core Doubao model is priced at only ¥0.0008 per thousand tokens—just 0.8 fen—for over 1,500 Chinese characters, 99.3% cheaper than industry standards. At this rate, one yuan buys 1.25 million tokens—approximately two million Chinese characters, equivalent to three copies of *Romance of the Three Kingdoms*. The 128K general-purpose Doubao model is priced at ¥0.005 per thousand tokens, 95.8% lower than market averages.
For comparison, GPT-4 Turbo charges $0.01 per 1,000 input tokens and ¥0.21 per 1,000 output tokens. By contrast, ByteDance has slashed prices so deeply it’s being dubbed the “Pinduoduo of AI.”
Doubao isn’t alone—many other domestic large models are also cutting prices.
Recently, Baidu released a lightweight version of its ERNIE model, with ERNIE Tiny priced at just ¥0.001 per thousand tokens—effectively ¥1 for one million tokens.
In May, Zhipu AI significantly reduced commercial pricing for its large models. The entry-level GLM-3 Turbo model saw an 80% price cut, dropping from ¥5 to ¥1 per million tokens, making it accessible to more businesses and individuals.
Zhipu AI’s large model pricing
On May 6, DeepSeek—a company under China’s prominent hedge fund High-Flyer Quant—launched its second-generation MoE large model, DeepSeek-V2. Its API pricing is set at ¥1 per million input tokens and ¥2 per million output tokens (with 32K context).
On May 9, Alibaba Cloud officially released Qwen 2.5. According to OpenCompass benchmarks, Qwen 2.5 matches GPT-4 Turbo in performance, while remaining freely available to individual users via app, website, and mini-program.
On May 14, Tencent open-sourced its Hunyuan text-to-image model for free commercial use.
Overseas, OpenAI’s recently launched GPT-4o has also seen steep price reductions—not only offered free to all users but also halving API costs compared to GPT-4 Turbo launched in November last year, while doubling speed. This marks the third major price cut for OpenAI’s large model products.
French AI firm Mistral AI’s Mistral Large model currently offers input and output pricing about 20% cheaper than GPT-4 Turbo, drawing significant attention.
Whether domestically or internationally, large models are collectively slashing prices.
Lowering Costs, Accelerating Application Deployment
Price wars among vendors are now fully underway. Yet just months ago, the prevailing wisdom was that training large models was extremely expensive. How then have companies managed to slash prices so dramatically in such a short time?
Tan Dai of Volcano Engine believes cost reduction is crucial for advancing large models into the “value creation stage.” For small and medium enterprises, cost is a primary consideration when adopting large models. Tan revealed that ByteDance has implemented numerous technical optimizations across model architecture, training, and production to achieve these lower prices.
OpenAI CEO Sam Altman takes pride in offering ChatGPT without ads: “One of our key missions is to provide AI products to people for free.”
Indeed, low prices help large model developers seize market opportunities and establish footholds. Increased user volume, in turn, helps refine and improve model training. But have training costs actually decreased?
When GPT-4 launched last year, Sam Altman disclosed that training OpenAI’s largest model cost “well over $50 million.” According to Stanford University’s *2024 AI Index Report*, GPT-4’s training cost is estimated at $78 million.
Such high training costs directly inflated usage fees, locking out many potential enterprise users.
However, researchers are actively exploring lower-cost training methods. Last year, researchers from National University of Singapore and Tsinghua University proposed VPGTrans—a framework enabling high-performance multimodal large models to be trained at extremely low cost. Compared to training vision modules from scratch, VPGTrans reduced BLIP-2 FlanT5-XXL training costs from over ¥19,000 to under ¥1,000.
Among domestic models, developers have found various ways to reduce costs and boost efficiency. After improving dataset quality and optimizing architecture, DeepSeek-V2 leveraged the AI heterogeneous computing platform “Baige” to increase throughput by up to 30% during training and 60% during inference.
Beyond training, even foundational infrastructure like chips are becoming cheaper. For example, price drops in NVIDIA’s A100 AI chips have reduced large model training costs by approximately 60%.
The most direct impact of the large model price war is accelerated application deployment. On the Doubao platform, over 8 million intelligent agents have already been created. The GPT Store hosts over 3 million apps built on GPT models.
In just six months, the era of competing solely on raw model performance through massive spending appears to be over. Today, with widespread price reductions, users increasingly prioritize which large model offers the best combination of affordability and performance—accelerating real-world adoption and commercialization of large model applications.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
