Written by: Rita

The rankings of large models are undergoing a reshuffle, with Anthropic taking the lead, while Meta and Google are still catching up.

In the cutting-edge AI data tracking report released by Bank of America on August 17, the latest data across five dimensions—model intelligence rankings, usage trends, token pricing, GPU rental costs, and memory costs—has been systematically sorted. The core judgment of the report is: the AI ecosystem is overall healthy, but the structure is undergoing subtle changes, with open-source models gaining in usage share, while the paid share remains highly concentrated in leading closed-source models; token pricing saw a surge in July but relaxed in August, while GPU rental costs remain robust.

Bank of America believes that the release of cutting-edge models and trends in token pricing will be key driving factors affecting the sentiment of large internet stocks.

Intelligence Rankings: Anthropic Leading, Meta and Google Catching Up

The three major benchmark tests point to the same conclusion: Anthropic is currently the leader in cutting-edge AI capabilities.

The top three in the Artificial Analysis Intelligence Index are dominated by Anthropic and OpenAI: Claude Opus 5 ranks first, Claude Fable 5 second, and GPT-5.6 Sol third. Meta MuseSpark 1.2 ranks seventh, while Google Gemini 3.6 Flash is only thirteenth. The Agency Index is also led by Claude Opus 5, followed by GPT-5.6 Sol and Qwen 3.8 Max in second and third place, respectively. Meta and xAI have made it into the top ten, but Google still trails behind. The top three in the programming agency index align with the intelligence index, with Grok 4.5 climbing to fifth place, and Meta temporarily at eleventh.

Bank of America points out that Meta and Google are still in a catching-up position; Meta's MuseSpark has entered the top ten, but there is still a gap to the leaders; Google's Gemini 3.6 Flash ranks outside the top ten across all lists.

Usage Distribution: Open-source Models Take Volume, Closed-source Models Take Revenue

The data on usage and payments shows a starkly different pattern.

Token usage data from the Vercel platform shows that DeepSeek leads with about 30% market share, followed by Anthropic at 25%, and OpenAI at 16%. However, in terms of revenue, Anthropic leads with about 65% share, while OpenAI accounts for only 11%, and Google for 8%.

Data from OpenRouter further confirms this divide: Chinese open-source models dominate in usage, with Xiaomi's MiMo-V2.5 ranking first with a monthly usage of 32.8 trillion tokens, followed closely by DeepSeek V4 Flash at 26.4 trillion tokens. In closed-source models, OpenAI's GPT-5.6 Luna leads with 8.6 trillion tokens, while Claude Opus is second with 4.8 trillion tokens.

Bank of America notes that the usage share of open-source models continued to grow in August, but the payment share remains highly concentrated in leading closed-source models like Anthropic and OpenAI. The pattern of “open-source taking volume, closed-source taking revenue” indicates that the commercialization capabilities of leading closed-source model vendors are far superior to those of open-source competitors.

Token Prices Relax, GPU Rental Costs Solid

The pricing data for August sends mixed signals.

The AI Token Price Index dropped by 9% month-on-month to $2.21 in August, but still rose by 87% year-on-year. The LLM Token Spending Index fell by 27% month-on-month to $1.15. Bank of America attributes the decline in prices mainly to OpenAI reducing the price of GPT-5.6 Luna by 80%, cutting the price of GPT-5.6 Terra by 20%, and Gemini 3.6 Flash replacing Gemini 3.5 Flash with a 17% cost efficiency advantage.

Bank of America indicates that token pricing was very strong in July, which may have helped the profitability of cloud service providers, but competition intensified in August. If this trend continues, the AI revenue growth of cloud providers may face pressure.

GPU rental prices remain solid. The rental price for B200 decreased by 2% month-on-month to $5.63 per hour, but increased 7% year-on-year; H100 rose by 2% month-on-month to $2.77 per hour, up 33% year-on-year; A100 remained steady at $1.65 per hour, up 17% year-on-year. Bank of America believes this conveys a clear message: the demand for AI infrastructure remains healthy.

DRAM and NAND prices remain high. DRAM (DDR5 24GB) rose by 8% month-on-month in August, and increased by 483% year-on-year; NAND (QLC 1TB) remained stable month-on-month, up 432% year-on-year. Bank of America notes that this continues to exert pressure on the profitability of cloud service providers.

Three Key Variables

The Bank of America report points to three key variables.

The first is the pace of model releases. Anthropic currently has an advantage in the competition for cutting-edge models, but the releases of Meta's “Watermelon” and Google's Gemini 4 could act as catalysts, potentially changing the current ranking landscape.

The second is the trend in token pricing. The surge in July helped cloud providers' profitability, but the price relaxation in August deserves attention. If price competition continues, cloud vendors’ AI revenue growth may face pressure.

The third is that the strength of GPU rental rates indicates that the lifespan of servers may be reaching or exceeding expectations. The A100 has been shipping since 2020 and is still being leased and used, which extends the depreciation cycle for GPU assets and supports long-term demand logic for suppliers like NVIDIA.

The core conclusion of the Bank of America report is: the demand for AI infrastructure remains healthy, but the competitive landscape is evolving. Leading closed-source models retain a significant lead in commercialization capabilities, but usage of open-source models is catching up; token pricing is beginning to relax after a surge, but GPU rents and memory prices have yet to signal any slowdown in demand.

Disclaimer

This article is a compilation and interpretation of a third-party brokerage research report (Bank of America Securities, August 17, 2026) by潮向研究, combined with publicly available market information. The ratings, target prices, earnings forecasts, and related judgments cited in the text are the opinions of the analysts of that brokerage and only represent the position of their respective institutions, not that of潮向研究, and do not constitute any investment advice.

The market carries risks, and decisions should be made independently. This article should not be used as the basis for buying or selling any securities.