Compared to open source models, cutting-edge models are like luxury handbags! Deutsche Bank: This may lead to the market reevaluating AI.

Deutsche Bank points directly to the most core but most underestimated structural contradiction in the current AI industry: there is a "cost chasm" between cutting-edge proprietary AI models and open-source/open-weight models that is staggering. On June 20, according to Wind Trade Platform, Deutsche Bank stated in its latest research report that the data is clear: Anthropic's top-tier cutting-edge model Claude Fable 5 scores 60 on the Artificial Analysis intelligence index, with an average weighted cost of about $3.25 per task; whereas DeepSeek V4-Pro scores 44, with a cost of only about 5 cents per task - the former is about 65 times the cost of the latter. However, for about 90% of daily ordinary tasks, the performance of the inexpensive models is almost equivalent to that of cutting-edge models. Deutsche Bank's core judgment is that the high premium of proprietary cutting-edge models is more like a "status symbol" pricing of luxury handbags, rather than a pure performance premium. As top AI companies prepare for IPOs and shift from a subscription model to token-based billing, corporate users will be forced to reconsider whether the "cutting-edge premium" is worth paying. This trend may trigger a more profound and long-lasting market reappraisal than the "DeepSeek moment" in early 2025. This may indicate that the "narrative of operating costs" in the AI industry is quietly replacing the "narrative of computing power demand" as a new pricing anchor. If the real cost-effectiveness of proprietary models is fully priced in by the market, the valuation pressure faced by AI-related stocks will be structural rather than temporary. How deep is the cost chasm between cutting-edge and open-source models? Deutsche Bank cited data from the Artificial Analysis intelligence index, comparing mainstream AI models on the market based on intelligence score and single task cost. Cutting-edge proprietary models: Anthropic Claude Fable 5 (with fallback mechanism): intelligence score of 60, average weighted cost of about $3.25 per task. OpenAI GPT-5.5 (super high configuration): in the high-cost range. Google Gemini 3.1 Pro Preview: also in the high-cost cutting-edge camp. Open-source/open-weight models: DeepSeek V4-Pro (maximum configuration): intelligence score of 44, cost of about 5 cents per task. Meta Muse Spark: in the low-cost range (price data not available). Nvidia Nemotron 3 Ultra: in the low-cost range. OpenAI gpt-oss-120b (high configuration): in the low-cost range. Deutsche Bank particularly points out that the low-cost camp is not solely made up of Chinese models. Meta, Nvidia, and OpenAI's own open-weight models are also in the low-cost range. Therefore, the essence of this competition is not "U.S. vs China", but "cutting-edge proprietary vs open-weight". Deutsche Bank uses a vivid metaphor to describe this misalignment of performance and cost: the cutting-edge model is a roaring new supercar, while the open-source model is a reliable second-hand family car. The report acknowledges that in the most difficult reasoning tasks and Agentic work, cutting-edge models do have real, significant performance advantages. The gap between an intelligence score of 60 and 44 is substantial when handling the most complex tasks. However, the key issue is that for about 90% of daily ordinary tasks, inexpensive open-source models can perform almost the same work at a cost that is only about 1.5% of the cutting-edge model's cost. This means that the premium that most enterprise users pay for cutting-edge models is not necessarily driven by actual business needs, but more by brand awareness, habitual inertia, and even some form of "identity identification with using the best AI" - which is highly similar to the psychological logic of consumers buying luxury handbags. The premium will not disappear, it will only "migrate" - but each layer is sliding towards zero. Deutsche Bank presents an important structural observation: the cost of AI capabilities is decreasing at a rate of about ten times per year, but the cutting-edge premium will not disappear as a result, it will continue to "migrate". The logic chain is as follows: today's cutting-edge models will become commoditized capabilities in the future; at the same time, a new generation of stronger cutting-edge models will emerge with a new high premium; therefore, the price gap between "best available" and "good enough" will structurally exist in the long term; but every point on the price curve is continuously sliding towards zero. This mechanism means that the premium for cutting-edge AI is a moving target, rather than a fixed moat. This requires continuous dynamic evaluation of pricing power and profit margins for AI companies, rather than a one-time judgment. Business model transformation under IPO pressure: from "subscription-based" to "usage-based billing". Deutsche Bank points out that the AI cost issue has become particularly urgent now due to a key business catalyst: top proprietary AI labs are preparing for IPOs, and the business model is shifting from fixed-rate subscription to usage-based pricing with tokens. This shift will directly pass on cost pressure to enterprise users. The report cites a highly convincing real case: Uber has already burned through its entire token budget in just four months, and now limits all employees to a monthly AI usage spending limit of $1500. Deutsche Bank believes that this case clearly demonstrates that when AI usage costs shift from "implicit subscription" to "explicit usage-based" pricing, enterprise cost consciousness will be rapidly activated. Users who only need a "reliable workhorse" rather than a "super sports car" will increasingly ask themselves seriously: is the cutting-edge premium really worth paying for? "The second DeepSeek moment": quieter, but potentially more profound. Deutsche Bank compares the current situation to the "DeepSeek moment" in early 2025 and makes a highly alert forward-looking judgment. Looking back at the "DeepSeek moment" in early 2025: the market realized that AI capabilities close to the cutting edge level could be built at a much lower cost than expected, causing a severe shock to AI stocks. However, due to the continuous rise in overall AI demand, the stock market was able to recover lost ground. Deutsche Bank's judgment is that the current brewing "narrative of operating costs" is the "quieter but more enduring sequel" to that shock. The core logic is that if proprietary AI models were previously priced and traded in part as "status goods" - that is, their high price was part of their appeal - then once their real cost-effectiveness is fully exposed and priced in by the market, it may bring about a second round of revaluation to the AI valuation system. This shock will not be as dramatic as before, but the impact will be more profound and long-lasting. Deutsche Bank leaves an intriguing open ending: unless, like luxury handbags, the high price of AI itself is its ultimate selling point. The report also cites research data from Epoch AI, providing independent corroboration for the analysis: the average gap between the U.S. and China in cutting-edge AI capabilities is about seven months; Epoch AI also points out that this gap aligns almost completely with the gap between proprietary models and open-weight models. This finding further strengthens Deutsche Bank's core argument: the "U.S.-China AI gap" on the geopolitical dimension and the "proprietary/open-source gap" on the business dimension are essentially two expressions of the same chasm. This means that when assessing AI geopolitical risks, they should not be viewed separately from the logic of business competition. This article is reprinted from "Wall Street View", author: Dong Jing; GMTEight editor: Yan Wencai.