Token Prices Collapse 50% in One Summer

The cost of artificial intelligence has hit a historic milestone, crossing below the one-dollar threshold for the first time ever. The LLM Token Expenditure Index, a widely watched benchmark that tracks the effective market price of large language model tokens, plunged to 97 cents per million tokens on Monday. This represents a dramatic decline of more than fifty percent from its peak earlier this summer, marking the lowest reading since the index was launched late last year.

The plummeting prices signal an aggressive price war sweeping through the AI industry, fundamentally reshaping the economic landscape for both providers and consumers. Developers and businesses deploying AI at scale stand to benefit significantly from cheaper inference costs, potentially making AI agents, coding assistants, customer service systems, and enterprise automation far more economical than ever before.

However, the story looks very different from the perspective of frontier AI labs. Companies like OpenAI and Anthropic, which have committed billions of dollars to building massive computing infrastructure, now face a troubling paradox. While their fixed costs remain unchanged, the amount they can charge per token continues to fall. This creates what industry analysts describe as token deflation, where revenue lines compress even as compute commitments stay rigid.

The primary driver behind this price collapse appears to be the rapid rise of lower-cost open-source models emerging from China. Moonshot AI's Kimi K3 and other open-weight alternatives are undercutting premium offerings from established players, pulling down prices across the entire market. In response, OpenAI recently cut prices across parts of its GPT-5.6 lineup, while other providers have introduced dynamic pricing structures that allow access rates to fluctuate with demand. These competitive pressures have only intensified the downward spiral.

Industry observers note that the strategic response from AI companies is already becoming visible. Charles-Henry Monchau, chief investment officer at Syz Group, points out that foundation model labs are the most directly exposed to this trend. The competitive moat must shift away from raw model capability, where the open-weight gap is now measured in mere months, toward distribution channels, persistent memory capabilities, context handling, and broader software ecosystems.

The timing could not be more delicate for OpenAI and Anthropic, both of which have reportedly submitted confidential IPO filings to regulators this summer. Lower token prices risk conditioning consumers to expect cheaper access, which can permanently erode pricing power for providers over time. This threatens not only their revenue models but also the prospects for recouping the massive AI investments made by technology giants like Nvidia and Microsoft.

Silicon Data's head of research, Steve Hou, suggests that the recent price declines may signal that existing supply across both frontier and lower-cost models is already sufficient to handle most tasks. If prices continue falling faster than AI demand grows, both model providers' revenue and returns on AI infrastructure investment could fall short of market expectations.

The ripple effects are already visible in financial markets. Technology shares were under significant pressure on Tuesday, with the Nasdaq Composite sliding nearly one percent and the Philadelphia Semiconductor Index dropping more than two percent. Growing concerns about the sustainability of AI investment weighed heavily on investor sentiment.

Anthropic has responded to these pressures by unveiling its latest frontier models, Mythos 5.1 and Fable 5.1, which deliver improved performance while requiring fewer tokens for the same tasks. This efficiency gain helps reduce cost burdens, though access to Mythos remains currently limited due to cybersecurity concerns.

As the AI industry navigates this new reality, one fundamental question looms larger than ever: whether soaring demand for artificial intelligence can grow quickly enough to compensate for the falling price of intelligence itself. The answer to that question may determine which companies survive the current price war and which become casualties of their own massive infrastructure bets.