• Open-weight models hit 56% of Vercel AI Gateway token volume in August, up from 7% in December 2025.
• Anthropic captured 64% of AI Gateway spend while open-weight models took only 14%.
• Vercel reported average token prices fell 23.2% month-over-month in August, a third straight decline.
Open-Weight Models Hit 56% of Gateway Tokens
Open-weight AI models processed 56% of all tokens crossing Vercel's AI Gateway in August 2026, according to the company's own AI Gateway Production Index published on September 17, yet they captured only 14% of spend. The index, built from anonymized and aggregated traffic passing through Vercel's gateway, exposes a widening gap between usage and cost in enterprise AI. As recently as December 2025, open-weight models held under 7% of token volume; an April reading of 13% stems from a methodology change rather than a surge, so the share moved from niche to majority in roughly eight months. The gateway itself is the routing layer that mediates requests between applications and a menu of AI models, and Vercel's token count includes input and output plus reasoning, cache-read and cache-write tokens — a measure of total model work, not just prompt traffic. August also logged a 23.2% month-over-month drop in average token prices, the third consecutive monthly decline. Teams that exceeded 10 million tokens in each of the past two months saw median costs fall 7.6%, a figure Vercel says uses a different denominator from the overall average and should not be conflated with it. Anthropic dominated the spend column at 64%. The company's own explanation is routing, not substitution: its models tend to be picked for higher-priced, high-difficulty work, while low-cost open-weight systems absorb bulk processing — cheap models for volume, frontier models for the tasks where quality or reasoning depth pays. Daily leaderboard data currently shows open-weight share at 78.4%, but that reflects a specific window across three months and is not directly comparable with the 56% August figure. Two limits apply: the dataset is one infrastructure provider's anonymous sample rather than the whole AI market, and teams running open weights in-house bear GPU, deployment and maintenance costs beyond the models themselves. Where each workload lands on that cost-versus-control curve is, in effect, the tokenomics question of the current AI cycle.
Anthropic Weighs a New Model as Astra Gains
Anthropic is weighing the launch of a new AI model in response to the fast enterprise uptake of OpenAI's GPT-6 Astra, according to three people familiar with the matter; the company declined to comment. The deliberation comes days after CEO Dario Amodei argued in a September 12 essay that the industry should slow the pace at which AI capabilities improve. Per the same sources, Anthropic is evaluating the next model's safety while balancing new-model investment against profitability, as rising rates make investors more sensitive to when returns materialize. OpenAI launched GPT-6 Astra on September 3, with upgrades aimed at computer use, software engineering, cybersecurity and professional work tasks. Enterprise spend platform Ramp puts Astra at roughly 13% of tracked corporate AI spend, against about 8% for Anthropic's Claude Fable. On the developer platform OpenRouter, spending on OpenAI models overtook Anthropic's last week — the first time in two and a half years OpenAI has led on that metric. Revenue context sharpens the contest: Anthropic's annualized revenue climbed from about $9 billion at the end of last year to more than $65 billion by late July, while OpenAI crossed $40 billion in July; Anthropic projects roughly $190-200 billion by 2028. Some investors still frame Astra's growth as manageable, given Anthropic's entrenched enterprise base. The bigger near-term variable is the IPO calendar: Anthropic's listing marketing had been expected to begin as early as mid-October, a timeline already pushed back once, and the company may now delay the offering until after the November midterm elections. Open-weight diffusion adds a second pressure: investors note that enterprises building their own AI infrastructure can cut token costs and reduce reliance on external providers, and Meta — one of Anthropic's major customers — is exploring ways to use its models less as it builds in-house capability. Meta did not respond to a request for comment.
The Compute-Cost Read for Bitcoin
Read together, the two datapoints describe an AI market splitting its bill: volume migrates to cheap open-weight weights while premium dollars stay concentrated in frontier providers — a repricing driven by routing, in Vercel's own accounting, not a wholesale substitution of closed models. That cost deflation and the drift toward self-hosted infrastructure are exactly the arbitrage pitched by decentralized AI networks such as Bittensor (TAO), and they help explain why AI capex and crypto risk appetite keep moving together. With Bitcoin (BTC) trading near $81,208 at publication, we read the shift as a cost tailwind for compute-heavy plays — from chip-equipment names like Applied Materials to copper demand from data-center buildouts.
