Frontier AI model capabilities, usage, pricing and infrastructure demand: BofA sees healthy AI infrastructure demand as model competition and usage shares continue to shift

Overview

This September 2026 BofA Frontier AI Tracker reviews model capability rankings, developer usage, inference pricing, GPU rentals and memory costs as indicators for large-cap Internet-stock sentiment and AI investment economics. The report finds healthy AI demand and increasingly competitive model performance, with Meta making progress but Anthropic and OpenAI still leading key capability measures.

Core views

BofA frames frontier-model capabilities, model adoption, token pricing and infrastructure costs as key inputs into AI-related Internet-stock sentiment and valuations. On capability, Artificial Analysis ranked Claude Fable 5.1 and GPT-6 Astra jointly first in the Intelligence Index at 53. Meta’s MuseSpark 1.3 ranked fifth at 48, while Gemini 3.8 Flash ranked 12th at 41. Claude Fable 5.1 also led the Agentic Index at 58; Meta’s MuseSpark 1.3 tied for second at 56 with Claude Opus 5. In coding-agent capability, Claude Fable 5.1 and GPT-6 Astra tied for first at 62, versus MuseSpark 1.3 in seventh at 54 and Gemini 3.8 Flash in 14th at 42. BofA interprets Meta’s latest results as evidence that Meta Superintelligence Labs is narrowing the gap to the frontier and remains constructive on its upcoming AI products and expected October Watermelon LLM launch. Usage data show a different competitive picture. On OpenRouter, DeepSeek V4 Flash 0731 led September MTD token volume at 50.7 trillion tokens, followed by GPT-5.6 Luna at 45.1 trillion and Hy4 preview at 34.6 trillion. Among closed models, GPT-5.6 Luna led at 45.1 trillion tokens, ahead of Ox Alpha at 27.2 trillion and Gemini 3.7 Flash at 8.92 trillion. On Vercel’s AI platform, DeepSeek’s average MTD token share rose to 45.7% from 30.1% in August, while Anthropic fell to 13.6% from 22.1% and OpenAI fell to 12.0% from 14.9%. Meta gained 1.9 percentage points. In spend, Anthropic remained first at 52.0%, although down from 63.9%; OpenAI rose to 21.8% from 13.3%, and Moonshot.ai rose to 8.2% from 4.5%. BofA emphasizes that OpenRouter and Vercel data indicate developer adoption and commercial traction on those gateways, not total industry usage, because direct provider APIs, cloud platforms, private deployments, self-hosted inference and internal hyperscaler workloads are excluded. Inference-cost measures gave mixed signals. The AI Token Price Index rose 5% month on month and 104% year on year to $2.41 as of September 10, following an August decline. BofA attributes the increase partly to GPT-6 Astra entering the index at $10 per million input tokens and $50 per million output tokens, matching Fable 5.1 as the highest-priced model. GLM-5.3-Flash entered at $0.15 input and $0.50 output per million tokens, setting a lower price floor, while DeepSeek V4.1 Flash replaced V4 Flash at a lower price. In contrast, Silicon Data’s LLM Token Expenditure Index, a broader usage-weighted measure of effective inference cost, declined 11% month on month to $0.98 as of September 13 from $1.10 in August. On cost per Intelligence Index task, Claude Fable 5 with fallback was highest at $8.75, Claude Fable 5.1 Max with fallback was $7.63 and Claude Opus 5 Max was $5.86; GPT-6 Astra Max was $3.26, which OpenAI characterized as compelling price performance relative to leading competitors. Infrastructure indicators continued to support BofA’s view of healthy AI demand. As of September 13, B200 rental prices were up 1% month on month and 9% year on year to $5.68 per GPU-hour. H100 rentals fell 4% month on month but remained 30% above the prior year at $2.65, while A100 rentals fell 3% month on month but were 15% higher year on year at $1.60. BofA notes that the data suggest server useful lives remain healthy, given that A100 shipments began in 2020. Memory pricing was largely stable sequentially but sharply higher year on year: DDR5-5600/6400 24GB DRAM rose 1% month on month and 476% year on year, while 1TB QLC NAND declined 1% month on month but rose 437% year on year. Overall, BofA concludes that token pricing has stabilized after August’s decline, closed-model leaders remain closely matched on intelligence, open models continue to gain usage share, and firm GPU rentals point to sustained infrastructure demand. The report also cautions that renewed debate around AI safety and regulation could influence near-term sentiment toward hyperscalers. For Meta specifically, BofA maintains Buy and links its constructive stance to improving model competitiveness and upcoming AI-product launches.

Analysis framework

BofA tracks the frontier AI value stack by comparing third-party model benchmarks, gateway token and spend shares, published and effective token-cost indices, GPU rental rates, and spot memory prices. It uses the combined trends to assess competitive positioning, AI inference economics and infrastructure-demand conditions relevant to Internet stocks.

Methodology notes

• Competition & strategyEconomic Moat and Competitive Advantage

  Comparative frontier-model capability rankings

  The report compares intelligence, agentic and coding-agent benchmark scores to assess relative model capability and whether Meta is closing the gap with the leading AI labs.

• Industry AnalysisSupply-demand framework

  AI infrastructure demand assessed through GPU rentals and memory prices

  GPU rental rates and memory-price movements are used as indicators of demand conditions and cost pressure across the AI infrastructure stack.

• Other

  Token Price Index and LLM Token Expenditure Index

  The report uses a provider-price index and a broader usage-weighted effective-cost benchmark to distinguish listed frontier-model pricing from what users effectively pay for inference.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Meta Platforms (META)

  BofA says Meta’s latest models are narrowing the frontier-intelligence gap and maintains Buy.

  Strengths

  MuseSpark 1.3 ranked fifth in the Intelligence Index and tied for second in the Agentic Index; BofA is constructive on upcoming AI products and the expected October Watermelon LLM launch.

  Weaknesses

  MuseSpark 1.3 ranked seventh in the Coding Agent Index, behind the leading Anthropic and OpenAI models.

  Comparison

  Claude Fable 5.1 and GPT-6 Astra led intelligence and coding-agent measures; Gemini 3.8 Flash ranked below MuseSpark 1.3 in the cited Intelligence Index.

  Risks

  BofA cites AI investment spending that could hurt margins, competitive pressure for engagement and advertising budgets, macro-sensitive advertising revenue, and regulatory overhang.

Key data

• Top Intelligence Index score53Claude Fable 5.1 and GPT-6 Astra were jointly highest according to Artificial Analysis.
• Meta MuseSpark 1.3 Intelligence Index48; fifth placeGemini 3.8 Flash ranked 12th at 41.
• DeepSeek Vercel token-volume share45.7% September MTDUp from 30.1% in August.
• Anthropic Vercel spend share52.0% September MTDDown from 63.9% in August, but still the highest spend share.
• AI Token Price Index$2.41Up 5% month on month and 104% year on year as of September 10.
• LLM Token Expenditure Index$0.98Down 11% month on month from $1.10 in August as of September 13.
• B200 GPU rental price$5.68 per GPU-hourUp 1% month on month and 9% year on year.
• DDR5 DRAM price change+1% month on month; +476% year on yearFor DDR5-5600/6400 24GB 3Gx8 as of September 13.

Impact & implications

BofA views firm GPU rentals and elevated memory prices as evidence that AI infrastructure demand remains healthy, even as model usage shifts toward lower-cost open offerings. It sees Meta’s improving benchmark placement as supportive of its AI-product outlook, while noting that AI safety and regulation could weigh on hyperscaler sentiment in the near term.

Risks

• AI safety and regulation debates may influence near-term hyperscaler sentiment.
• Meta faces risks from macro-sensitive digital-advertising revenue, elevated AI spending, fixed-asset growth, AI-native competition and regulatory or litigation outcomes.

What to watch

• The expected October launch of Meta’s Watermelon LLM and related AI products.
• Whether Meta’s model rankings continue to narrow the gap with frontier leaders.
• Token-price and effective-inference-cost trends after September’s divergent readings.
• GPU rental rates and memory costs as indicators of AI infrastructure demand.
• Further changes in open-model usage share and AI safety or regulatory headlines.