In September 2026, the Silicon Data LLM Token Expenditure Index hit $0.97 per million tokens. It was the first time the usage-weighted average across more than 200 models dipped below the dollar mark since the index began tracking. The era of intelligence as a scarce, high-margin commodity is effectively over.
The Commodity Floor
The collapse is driven by a relentless, global race to the bottom. Open-weight models, particularly those originating from Chinese labs, now account for approximately 61% of top-model token traffic on OpenRouter. With the average cost of open-weight models sitting at $0.83 per million tokens-compared to $6.03 for proprietary alternatives-the market has spoken. Open-weight solutions are now 7.3 times cheaper, delivering 86% savings for developers who no longer require the absolute bleeding edge of reasoning.
Commoditization is a direct consequence of capability convergence. According to the Stanford HAI 2026 AI Index Report, the top four closed-source labs-Anthropic, xAI, Google, and OpenAI-were clustered within 25 Elo points on the Arena Leaderboard as of March 2026. When performance parity is reached, the model itself ceases to be a durable moat. We saw this play out in the frantic shipping window between August 22 and September 1, when OpenAI, Moonshot, Alibaba, and Z.ai all pushed frontier updates. The market is saturated with “good enough” intelligence, forcing a brutal reckoning for any business model built solely on selling raw inference.
The Frontier Ceiling
While the commodity floor crumbles, the frontier is moving in the opposite direction. Labs are not just maintaining prices; they are aggressively gating access and inflating costs for premium, autonomous-capable models. When GPT-6 Astra and Claude Fable 5.1 launched, they arrived at $10 per million input tokens and $50 per million output tokens-effectively doubling the cost of their predecessors.
This is the bifurcation we have been tracking. As we explored in our analysis of the $500-per-month agentic tier, the industry is pivoting away from selling tokens and toward selling proprietary workflow integration. The goal is no longer to provide the cheapest compute, but to capture the highest-value autonomous tasks. Even as OpenAI introduced the more accessible GPT-6.1 Sol at $2/$10, the strategic intent remains clear: keep the commodity tier cheap to maintain market share, while extracting massive premiums from enterprise users who need agentic AI that can actually execute complex, multi-step workflows.
Structural Realities
On one side, Gartner predicts a 90% reduction in inference costs for trillion-parameter models by 2030. On the other, the capital expenditure required to reach that future is staggering. Anthropic’s S-1 filing, as reported by Reuters, reveals $518 billion in cloud and compute commitments, with roughly 80% of that figure locked into binding, non-cancelable contracts. For every dollar earned in 2025, the company committed approximately $113 to future infrastructure.
This is the price of intelligence. The labs are betting that the massive, non-cancelable debt they are accruing today will be justified by the proprietary ecosystems they are building tomorrow. It is a high-stakes gamble on the transition from “model-as-a-service” to “infrastructure-as-a-moat.” Goldman Sachs notes that inference is already approaching 10% of total headcount costs at some software firms; as these costs scale, the pressure to move toward the $500-per-month agentic platforms will only intensify.
What Remains Unresolved
The “intelligence-as-a-utility” phase is collapsing. The market is bifurcating into a race for the bottom in raw compute and a race for the top in proprietary, high-cost agentic platforms. OpenAI’s revenue run rate of approximately $68 billion and its early discussions for a $30 billion funding round at a $1.4 trillion valuation suggest that investors are still buying the dream of the latter.
However, the convergence data from the Stanford HAI report suggests that the “frontier” is a moving target that is becoming increasingly expensive to hit. As we noted in our coverage of the GPT-6.1 Sol pricing shift, the labs are caught in a trap: they must continue to spend billions to stay ahead, even as the market for their core product-the token-is being hollowed out by cheaper, open-weight alternatives. The coming year will be defined by which firms survive the transition from selling intelligence to selling the workflows that make it useful.
Note: The Silicon Data LLM Token Expenditure Index (Bloomberg ticker SDLLMTK) is a usage-weighted average across 200+ models. Silicon Data raised a $30.5M Series A in August 2026. Stanford HAI Arena Leaderboard Elo figures are as of March 2026 and may have shifted since. Anthropic S-1 figures are from the prospectus as reported by Reuters – the filing has not yet appeared on SEC EDGAR. OpenAI revenue and valuation figures are self-reported and unconfirmed.