The cost of querying an artificial intelligence (AI) model performing at GPT-3.5’s level fell from $20 per million tokens in November 2022 to $0.07 per million tokens by October 2024, an over-280-fold reduction in about two years, according to Stanford’s AI Index 2025 Report. Tokens are the small chunks of text an AI model reads and produces, usually a few letters or a short word each. Companies price their models by the million tokens processed. OpenAI, Google and Anthropic have all cut or frozen those prices in the past six weeks.
Venture capital firm Andreessen Horowitz has tracked the same phenomenon since 2024, coining the term “LLMflation” to describe a roughly 10-fold price decline every year for a model of fixed capability, according to its analysis. That trajectory looks familiar from broadband internet: as infrastructure improves and competition intensifies, the price of access keeps falling while what a user can do with it keeps expanding. What changed this summer is the pace.
The three largest AI labs all moved in the same direction within weeks of each other this summer. OpenAI cut the price of GPT-5.6 Luna, its fastest and cheapest model, by 80%, from $1 to $0.20 per million input tokens. It cut GPT-5.6 Terra, its mid-tier model, by 20%, from $2.50 to $2, CNBC reported. Input tokens are what a company sends into a model. Output tokens, what the model sends back, run five to six times higher across all three labs. OpenAI left its flagship model, GPT-5.6 Sol, unchanged at $5 per million input tokens.
Google Cut Its Gemini Flash Price Weeks After Launch
Google arrived at a similar place from a different starting point. Gemini 3.6 Flash, a mid-tier model built to balance cost and capability, launched July 21 at $1.50 per million input tokens, and Gemini 3.5 Flash-Lite, Google’s cheapest and fastest model, launched the same day at $0.30 per million input tokens, according to Google’s pricing page.
Google then cut Gemini 3.6 Flash’s price to an introductory $0.75 within weeks. It priced its newest model, Gemini 3.7 Flash, which launched Thursday (Aug. 13), at that same introductory $0.75 per million input tokens through the end of the year, VentureBeat reported. Both rates double to $1.50 on Jan. 1, according to Google’s pricing page. That means the newest model does not undercut its immediate predecessor. It undercuts only Flash’s original launch price.
Anthropic Canceled a Price Increase It Had Already Announced
Anthropic went further. It reversed a price increase the company had already scheduled and publicly announced. Claude Sonnet 5 launched June 30 at $2 per million input tokens, explicitly framed as a temporary introductory rate, with a posted increase to $3 scheduled for Sept. 1.
On Aug. 10, roughly three weeks before that increase was set to take effect, Anthropic added an editor’s note to that same launch announcement canceling the increase and making the $2 rate permanent. The reversal took the only scheduled price increase among the three labs off the table.
The Same Pattern That Drove Broadband Adoption Is Now Driving AI
The pattern across all three companies is consistent: prices keep falling at the low and middle tiers even as flagship, most-capable models hold steady or fall more slowly. That mirrors broadband internet, where the cost of a baseline connection fell for years while the fastest, most premium tiers commanded a persistent premium. Falling prices at the bottom of the market did more to expand who could use the technology at all than any improvement to the top tier did. Enterprise AI workloads now range from simple jobs that barely tax a model, like sorting emails into categories, to multistep reasoning. The price gap between the cheapest and most expensive tiers has widened even as absolute prices fall across the board. At OpenAI, the spread between its cheapest and most expensive models widened to 25-to-1 from 5-to-1 in a single announcement.
That gap is what is letting AI spread into uses that were not affordable a year ago. A task that costs a fraction of a cent to run through a discounted, high-volume model can now be applied at a scale that would have been too expensive at 2023 pricing. It is the same dynamic that allowed cheaper broadband to unlock video streaming and cloud computing once connectivity costs fell far enough.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
