AI token prices hit record low as index falls below $1

The cost of running artificial intelligence models has fallen to its lowest point on record, capping a months-long price war led by Chinese AI labs that has cut inference costs by more than half since early summer.

Silicon Data's LLM Token Expenditure Index — a daily benchmark tracking the usage-weighted price of one million large language model tokens — dropped to 97 cents on Monday, according to CNBC. That marks the index's lowest reading since its creation in late 2025 and represents a decline of more than 50% from its summer high. The index had nearly doubled between its December 2025 inception and a peak in May, when it reached roughly $2.12, before reversing course as aggressive pricing from Chinese competitors reshaped the market.cnbc+1

A Price War With No Clear Floor

The collapse in token prices traces back to a wave of cuts that began in June, when five Chinese AI laboratories — including DeepSeek, Alibaba ↗ 1.28% Cloud, ByteDance, Moonshot AI, and Baidu ↗ 3.01% — simultaneously slashed their international API prices by 50 to 99 percent. DeepSeek's V4-Flash model, released in early August, priced input tokens at roughly $0.14 per million — about 1% of what Anthropic charges for comparable output. The move forced DeepSeek to abandon a planned peak-hour surcharge it had announced just weeks earlier.linkedin+3

By early August, average inference prices had fallen to a range of $1.16 to $1.18 per million tokens, according to Jefferies, citing Silicon Data's index — already the lowest level of the year at that point. The decline continued through August, pushing the index below a dollar for the first time.scmp

Spending Up Even as Prices Drop

The plunge in unit costs has not translated into lower overall AI spending. Token prices have collapsed more than 90% since 2023, yet total expenditure has roughly doubled since last year, according to Bloomberg. Chinese banks illustrate the pattern: several disclosed token consumption figures for the first time in their 2026 semi-annual reports, with one city commercial bank reporting an 18-fold increase in daily average token usage compared with the end of 2025. China Merchants Bank disclosed daily average consumption of 33 billion tokens.moomoo+1

The divergence between falling prices and rising spend has drawn comparisons to past technology cycles. Allianz Research has pegged the AI capital expenditure-to-sales growth gap at roughly 46%, exceeding the 32% divergence seen during the 2001 telecom excess cycle. Wang Xiaoqing, president of China Merchants Bank, told an August 31 performance briefing that the era of intelligence "may not necessarily be directly reflected in financial statement earnings, and may even manifest as expenses and costs in certain stages".aiweekly+1