DeepSeek has released the V4.1 Flash-1017 large language model. While the market was generally anticipating updates from domestic AI large models, DeepSeek took the lead in launching a new version, bringing substantial performance improvements with this upgrade.
Specific performance levels and changes were reported previously. You may have noticed that DeepSeek also released a Pareto curve this time, commonly known as the "AI kill zone." V4.1 Flash-1017 has once again reached god-like levels, with per-task costs only 1/50 to 1/100 of those of large models like Fable and Astra, offering at least 40-50 times better cost-performance ratio.
How did DeepSeek achieve such a dominant position? Everyone knows they have extremely strong AI optimization capabilities and have gone to great lengths to reduce costs. The chart below, released by the vllm_project, clearly demonstrates how powerful DeepSeek's optimizations on AI hardware are.
This chart compares results from the SemiAnalysis AgentX benchmark. On the GB300 NVL72 and Dynamo vLLM platforms, V4.1 Flash achieved a result of 1.3 billion tokens per dollar, with P90 interactive speed exceeding 100 tok/s/user.
At $1 for 1.3 billion tokens, the cost per 1M tokens is just $0.00077, which is far cheaper than V4.1 Flash's input price of $0.15.
Of course, these two prices cannot be directly equated so simply. The TCO cost of $1 for 1.3 billion tokens refers to the compute side. API pricing involves more complex considerations. However, DeepSeek's ultra-low costs give them an unbeatable advantage in this regard.
Precisely because DeepSeek V4.1 Flash has such an overwhelmingly advantageous cost structure, major platforms are bold enough to offer ultra-low pricing. In the $10 packages for OpenCode and CommandCode, recent usage values have reached $80-$120 worth of tokens. OpenCode even went free recently. On aggregate routing platforms, V4.1 Flash's own usage is now at the hundred-trillion-token level, equivalent to the combined usage of six or seven competitors.
We used AI to compare the TCO of multiple open-source large models. For models with up to one trillion parameters, their output is at most 1/10th that of V4.1 Flash. For large models with 1.5-3 trillion parameters, the cost difference can reach 40-50 times or more, making them completely incomparable.
It should also be noted that this TCO figure is for V4.1 Flash-0910, not the version 1017 released today. Although the API pricing for the latter remains unchanged, it has significantly optimized token efficiency, reducing token consumption. The extent varies by task, but subjectively, consumption has dropped by at least half, meaning the TCO could see another significant decrease.