- Published on
After DeepSeek V4 Flash, large models are starting to compete on 'intelligence-efficiency ratio'
- Authors
- Name
- DeepHorus
After the release of DeepSeek V4 Flash, the large model industry has shifted its focus from pure "intelligence ceiling" to "cost-effectiveness ratio"—that is, the ability of a model to solve real-world problems per unit of cost. Behind this trend is the explosive growth in API calls driven by agent-based applications, along with the resulting high computing bills. Tech giants like Microsoft have even faced significant cost pressure due to employees overusing AI, forcing the industry to rethink the economics of models.
Previously, the industry competed on scores in various benchmark tests, such as SOTA performance, while ignoring token consumption in actual deployment. However, when agents need to autonomously search, read files, write code, and iterate repeatedly, a single task can trigger hundreds of API calls. For example, data from OpenAI in May 2026 shows that over 70% of users submitted Codex tasks that would require at least an hour of human work, and the most active 1% of users generated more than 60 hours of agent runtime in a single day. This high-frequency usage has made "cost per task" a key metric.
Against this backdrop, the APPSO editorial team designed a series of real-world tests, giving different models a one-dollar budget to complete the same tasks in order to evaluate their cost-effectiveness ratio. The first task was to build an unofficial status monitoring page for the DeepSeek API, requiring the model to independently research information and design an original mascot. DeepSeek V4 Flash Max called the model 25 times, processed 1.22 million input tokens and 67,000 output tokens, and cost only $0.0758 in total. Subsequently, testers had the model generate a cinema viewing guide for "The Odyssey." Although the result was detailed, the style was lacking; switching to Claude Sonnet 4.6 produced excellent results, but the cost soared to $2.5.
In search of a cost-effective alternative, APPSO discovered Ant Group's Ling-3.0-Flash model. It scored 38 points on the intelligence index from third-party evaluation agency Artificial Analysis, tying with MiMo-V2.5 and Qwen3.6 27B, but it activates only 5.1B parameters during inference, significantly lower than the 10B-15B active parameter scale of similar models. In tests, Ling-3.0-Flash completed the status monitoring page task for just $0.0402, 40% cheaper than DeepSeek; on the cinema guide task, it took 17 minutes and 55 seconds and consumed $0.483, only one-sixth the cost of Claude Sonnet 4.6. Although it recommended the IMAX 70mm format, which cannot be viewed in mainland China, the low price means multiple retries are still more cost-effective than expensive models.
Data shows that Ling-3.0-Flash is particularly well-suited for batch extraction tasks with short inputs and fixed fields, as well as agent scenarios requiring high-frequency API calls. The larger the call volume, the more obvious its cost advantage. Hugging Face co-founder Clem pointed out that the cost per task varies by about 800 times across different models: flagship models average over $31 per task, while DeepSeek V4 Flash Max costs only $0.04. Open-source models have thus become the cost-effective choice. Data from the OpenCode platform shows that the official DeepSeek V4 Flash has consumed 8 trillion tokens through the platform, with daily consumption surpassing the average daily scale of the entire OpenRouter platform.
The "cost-effectiveness ratio" is not simply a price war; it requires models to deliver stable output with fewer tokens, less time, and less computing power while ensuring task completion. DeepSeek V4 Flash covers a wide range of tasks with a 1M context and low price, while Ling-3.0-Flash uses 5.1B activated parameters to achieve high throughput and fast response. The two represent two directions on the high efficiency route and are reshaping the competitive logic of large models—no longer blindly stacking parameters, but truly integrating intelligence into business processes.
For ordinary users, when each API call is no longer so expensive that it requires careful budgeting, AI can transform from an impressive demo into reliable productivity. As industry insiders say, AGI remains the goal at the mountaintop, but the vast number of daily tasks at the foot of the mountain—reading code, researching information, running workflows—need economically efficient intelligence even more. This competition over "cost-effectiveness ratio" may well determine whether AI can truly be scaled and implemented.
