- Published on
What Exactly Is It Cutting?
- Authors
- Name
- DeepHorus
Recently, the open-source AI Agent tool OpenCode released data showing that the official version of DeepSeek V4 Flash consumed 800 billion tokens in a single day on the platform, with 500 billion from free allowances and 300 billion from paid plans. In comparison, OpenRouter, which integrates over 400 models, processes approximately 6.6 trillion tokens per day on average. This means that the usage of a single model at a single entry point has exceeded the entire network scale of a large routing platform.
Behind this phenomenon lies a fundamental shift in the way AI is used. In the past, users asked questions and received answers, but today's Agents autonomously read files, modify code, run programs, and iterate repeatedly. A single task can last for hours and invoke dozens of tool calls. For example, when Claude Opus 5 was used to create a single-page 3D game, the model continuously generated, took screenshots to check, modified, and eventually produced just one HTML file, consuming 690 million tokens at a cost of nearly 3,000 RMB.
DeepSeek V4 Flash continues the low-price strategy, charging 2 RMB per million output tokens, while Anthropic's Claude Opus 4.8 standard price is $25 (about 170 RMB), an approximately 85-fold difference. In the LMArena frontend coding evaluation, V4 Flash's preliminary ranking temporarily places it ahead of Opus 4.8 Thinking. In Artificial Analysis' Intelligence Index v4.1, V4 Flash Max scores 50, while Claude Opus 4.8 Max scores 56, but the API call costs to run the same evaluation suite are $72.02 and $3752.55 respectively, with the latter being about 52 times higher.
This combination of "slightly better performance but dozens of times higher price" is reshaping model selection logic. The weakest small models are still suitable for simple tasks, and the strongest flagship models are used for difficult scenarios. The middle-tier models, however, face an awkward position: they are only slightly better than V4 Flash but cost dozens of times more, yet they are not strong enough to reliably solve tasks that V4 Flash cannot. This is called the "DeepSeek slash line"—by virtue of a nearly two-order-of-magnitude price difference, it pushes more expensive models out of the default invocation position.
V4 Flash's low cost stems from architectural innovation. It has a total of 284 billion parameters, but only about 13 billion are activated per token, using mixed attention, low-precision weights, and cache optimization. According to DeepSeek's publicly available technical report, under a million-token context, V4 Flash's per-token inference computation is about 10% of V3.2's, and KV Cache is about 7%. V4 Pro has a total of 1.6 trillion parameters with about 49 billion activated. Post-training focuses on strengthening coding and Agent capabilities, with a dedicated tool-calling format designed, and using hundreds of thousands of concurrent sandboxes to let the model practice running code, reading error messages, and continuing to modify.
The official version of V4 Flash has the same structure as the Preview, with only re-post-training, scoring 82.7 on Terminal Bench 2.1. The community project DwarfStar, through low-bit quantization, fits the q2 version into 96GB or 128GB unified memory devices, achieving a generation speed of over twenty tokens per second on some 128GB Macs. This means that cutting-edge Agent models are gradually moving toward local devices; although the full weights still require about 167GB of storage, quantized deployment has opened new possibilities.
