The GitHub trending page and Hugging Face download charts shifted violently this week. A sudden surge of activity signaled what many in the developer community are already calling the second DeepSeek moment. Across X and Discord, engineers are not just discussing a new release but are actively circulating benchmark screenshots that suggest the economic barrier to high-tier AI has just collapsed. The catalyst is the arrival of DeepSeek V4, a release that transforms the conversation from how much performance we can buy to how much performance we can get for nearly nothing.
The Architecture and Economics of DeepSeek V4
DeepSeek, the Chinese AI startup that previously disrupted the industry with its R1 model in January, has now deployed V4. This latest iteration is a Mixture-of-Experts (MoE) model boasting 1.6 trillion parameters. Unlike many of its frontier competitors, DeepSeek V4 is released under the MIT license, granting developers total freedom for commercial use, modification, and redistribution. The model is available for immediate download via Hugging Face or through a dedicated API. Deli Chen, a researcher at DeepSeek, noted on X that the release was the culmination of 484 days of effort, framing the project as a step toward making AGI accessible to everyone.
The pricing structure for the Pro model is designed to aggressively undercut the market. For cache miss scenarios, where the model cannot reuse previous computations, the cost is $1.74 per million input tokens and $3.48 per million output tokens. A standard request involving one million tokens of each results in a total cost of $5.22. When a cache hit occurs, the input price drops significantly to $0.145 per million tokens, bringing the total for the same volume down to $3.625.
bash
DeepSeek V4 API 호출 예시 (curl)
curl https://api.deepseek.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $DEEPSEEK_API_KEY" \
-d '{
"model": "deepseek-v4-pro",
"messages": [{"role": "user", "content": "Hello"}]
}'
For those prioritizing speed and cost over maximum reasoning depth, the Flash model offers an even more drastic reduction. In cache miss scenarios, the input cost is $0.14 per million tokens and the output is $0.28 per million, totaling $0.42. With a cache hit, this total drops further to $0.308. While the Flash model does not match the Pro version in raw capability, its pricing is more than 98% lower than that of GPT-5.5 or Claude Opus 4.7.
Breaking the Performance-Price Correlation
For the past several years, the AI industry operated under a rigid axiom: elite performance required an elite budget. To access the highest reasoning capabilities, enterprises had to accept high API overhead. OpenAI's GPT-5.5 exemplifies this trend, charging $5 per million input tokens and $30 per million output tokens, totaling $35 for a million-token pair. Anthropic's Claude Opus 4.7 follows a similar pattern with $5 for input and $25 for output, totaling $30.
DeepSeek V4 Pro shatters this correlation. At $5.22 for a cache miss, it costs approximately 1/7 the price of GPT-5.5 and 1/6 the price of Claude Opus 4.7. When cache hits are factored in, the gap widens further, with DeepSeek V4 costing roughly 1/10 of GPT-5.5 and 1/8 of Claude Opus 4.7. This is not merely a discount; it is a fundamental shift in the economics of inference. Tasks that were previously deemed too expensive to automate at scale now suddenly possess a viable business case.
This economic shift is amplified by the MIT license. The ability to download the model and host it on private infrastructure or perform specialized fine-tuning removes the dependency on a third-party API provider. For startups and independent developers, the combination of near-frontier performance and zero-cost licensing creates a strategic advantage that was previously reserved for companies with massive compute budgets.
However, the performance data reveals a nuanced picture. In the BrowseComp benchmark, which measures web searching and navigation capabilities, DeepSeek V4 Pro Max scored 83.4%. This puts it ahead of Claude Opus 4.7 at 79.3% and within a slim 1 percentage point margin of GPT-5.5's 84.4%. While it still trails the GPT-5.5 Pro version's 90.1%, the result is startling for an open-weight model. This specific strength is critical because web browsing is the backbone of agentic AI, where models must independently navigate the internet to execute complex tasks.
Other benchmarks show that a gap still exists in deep academic and engineering reasoning. On the GPQA Diamond benchmark, DeepSeek V4 scored 90.1%, trailing GPT-5.5 at 93.6% and Claude Opus 4.7 at 94.2%. In the Humanity's Last Exam test without tool use, DeepSeek V4 recorded 37.7%, falling behind GPT-5.5's 41.4% and Claude Opus 4.7's 46.9%. Similarly, on the SWE-Bench Pro software engineering benchmark, DeepSeek V4 reached 55.4%, while GPT-5.5 and Claude Opus 4.7 scored 58.6% and 64.3% respectively.
The developer community is currently split between those who argue that GPT-5.5 still holds the crown for absolute reasoning and those who believe the price-to-performance ratio of DeepSeek V4 makes the difference negligible. The reality is that while DeepSeek V4 may not completely replace the absolute ceiling of proprietary models, it has effectively lowered the floor for what constitutes a high-performance AI pipeline.
DeepSeek V4 has successfully decoupled frontier-level intelligence from prohibitive pricing, forcing a total recalibration of how the industry values AI performance.
