美银前沿AI数据追踪器：模型排名、使用数据及Token/GPU成本

BofA Frontier AIDataTracker: Modelrankings, usage data, and token/GPU costs Industry Overview 17 August 2026 Key signals shaping Internet AI stocks sentimentFrontier AI capabilities& pace of AI investments remain key drivers of sentiment and valuation across large-cap Internet stocks. We are launching the BofA Frontier AI Trackerthat monitors key indicators across the AI value stack, including frontier modelintelligence rankings, usage trends, token pricing, GPU rental rates & memory costs. EquityUnited States Justin PostResearch AnalystBofAS+1 415 676 3547justin.post@bofa.com Intelligence: Claude leading, followed by GPT 5.6 SolModel intelligence remains a key driver of competitive positioning& AI sentiment. According to Artificial Analysis, Anthropic currently leads the Intelligence Index withClaude Opus 5 ranking first, followed by Claude Fable 5 and OpenAI's GPT-5.6 Sol.Meta’s MuseSpark 1.2 (xhigh) was ranked 7thand Gemini 3.6 Flash was ranked 13th,suggesting Meta (Buy) is making progress with its AI labs initiatives. Nitin Bansal, CFAResearch AnalystBofAS+1 415 676 3551nbansal7@bofa.com Model usage: A mixed bag based on platform end marketVerceldata suggests DeepSeek gained usage share m/m and leads at 30% share, while Anthropic usage share was second (25%) and was first for spend (65%), and OpenAIusage share was at 16% and spend share at 11%. OpenRouter data suggests openChinese models are leading for usage, with Xiaomi’s MiMo-V2.5 #1 at 32.8tr tokensMTD, followed by DeepSeek V4 Flash at 26.4tr. For closed models, OpenAI’s GPT-5.6Luna lead for usage with 8.6tr tokens, followed by Claude Opus 4.8 at 4.8tr. We notethis data does not incorporate model usage directly on leading hyperscaler platforms. MTD: Month to DateGPU: Graphic Processing UnitAPI: Application Programming InterfaceTPI: Token price IndexDRAM: Dynamic random-access memory Token pricingfalls m/m after 3 month run up As of August13, AI Token price Index (TPI) was down 9% m/m (still up 87% y/y) to $2.21and LLM Token Expenditure Index was down 27% m/m to $1.15. OpenAI price cuts,including an 80% reduction for GPT-5.6 Luna and a 20% reduction for GPT-5.6 Terra, aswell as Gemini 3.6 Flash replacing Gemini 3.5 Flash in the TPI index (reportedlydelivering 17% better cost efficiency) likely contributed to m/m declines. GPU rentals and memory pricing mixed m/mAccording to Silicon Data,B200 rental prices were down 2% m/m (+7% y/y) to $5.63 in Aug, H100 were up 2% m/m (+33% y/y) to $2.77 and A100 were flat m/m (+17% y/y) to$1.65. In Aug, DRAM memory prices (DDR5–5600/6400 24GB 3Gx8) were up 8% m/m(+483% y/y) and NAND memory prices (QLC Flash 1TB) were up flat m/m (+432% y/y).Data suggests server useful lives remain healthy (A100 shipments started in 2020). Healthy ecosystem, token costsworth watching We think August data suggests: 1)Anthropic and OpenAI are setting the pace at the LLMfrontier (Meta’s Watermelon and Google’s Gemini 4 launches are potential catalysts), 2)Lower-cost open models had modest recent usage share gains (Which may haveprompted Meta’s launch of open-weight models), 3) Token pricing was very strongthrough July, which likely aided cloud margins, but price competition increased in August,4) GPU rental pricing remains firm, suggesting AI infra demand remains healthy & serveruseful lives are likely meeting/exceeding expectations, and 5) no NT relief from memorycost inflation. We think hyperscaler model launches and token pricing trends andannouncements could be a key mega cap Internet sentiment drivers going forward. BofA Securities does and seeks to do business with issuers covered in its researchreports. As a result, investors should beaware that the firm may have a conflict ofinterest that could affect the objectivity of this report. Investors should consider thisreport as only a single factor in making their investment decision.Refer to important disclosures on page 12 to 16. Analyst Certification on page 11. PriceObjective Basis/Risk on page 11.13009068 Timestamp: 17 August 2026 04:30AM EDT BofA Frontier AI Tracker - Aug Overview AI Index Dashboard Frontier Model Intelligence Ranking Intelligence Index According to Artificial Analysis, Claude Opus 5 (max) was ranked highest in intelligenceindex, followed by Claude Fable 5 (w/fallback), and GPT-5.6 Sol (max). Meta’s MuseSpark1.2 (xhigh) was ranked 7thand Gemini 3.6 Flash was ranked 13th. Agentic IndexAccording to Artificial Analysis, Claude Opus 5 (max) was ranked highest in Agentic index, followed by GPT-5.6 Sol (max) and Qwen 3.8 Max. Meta’s MuseSpark 1.2 (xhigh)was ranked 9th, Grok 4.5 (high) was ranked 10thand Gemini 3.6 Flash was ranked 14th. Coding Agent IndexAccording to Artificial Analysis, Claude Opus 5 (xhigh) was ranked highest in Coding Agent index, followed by GPT 5.6 (Sol max), and Claude Fable 5 (max). Grok 4.5 (high)was ranked 5thand Meta’s MuseSpark 1.1 (xhigh) was ranked 11th. Coding Agent Benchmark-Top 15Models Top Models by Inte