AI Costs Are Falling, but Subscription Prices Aren't

AI has gotten cheaper, but subscription prices haven't. The cost of using AI at the same level of performance is falling rapidly, yet subscription fees at major AI companies have not come down. We take a look at why user prices are holding steady even as costs decline.

According to an analysis by AI research institute Epoch AI, the minimum cost of achieving a given level of performance has fallen by roughly 47% per quarter since 2023. Google and OpenAI, however, are restricting usage and model access while keeping subscription prices unchanged.

What Epoch AI measured is not a decline in all subscription prices. The report's cost figures refer to the price of the cheapest way to reach a target score on a specific benchmark, compared by switching models or adjusting compute (the amount of work a computer processes). The report made clear that this does not mean the costs AI companies actually spend to run their services are falling at the same rate. Epoch AI explained that technologies making AI cheaper are advancing rapidly. Still, the report's key point is that these cost declines are not immediately translating into subscription price cuts.

The cost of achieving a 75% accuracy rate on the GPQA Diamond test, which covers graduate-level science questions, was $0.30 per question for ChatGPT-o3. GPT-5.6 Luna, released about a year and a half later, achieved the same score at $0.0004 per question. Per 10,000 questions, that's a drop from $3,000 to $4—roughly a 725-fold reduction in the money needed to secure the same level of performance.

OpenAI has reduced the usage it provides to new subscribers of its ChatGPT Pro plan, which costs $200 a month. The reduced usage will also apply to existing subscribers who received a grace period starting on the 29th, with the subscription fee staying at $200 a month. Google will restrict access to its higher-tier Gemini Pro model on its Google AI Plus plan, priced at $4.99 a month, starting on the 9th. In this way, while the cost of obtaining the same level of AI performance is falling rapidly, the price consumers pay remains the same—or is becoming even less favorable.

That said, model operating costs and subscription prices are different concepts. The level of performance consumers demand from AI is also rising. Consumers now expect AI to go beyond simple queries to writing complex code, analyzing multiple documents, and performing long-running tasks. A customer who asks ten questions a day and one who asks hundreds generate the same revenue but incur different costs. Even as model efficiency improves, if usage grows faster than those gains, the savings are offset. In the end, the balance between 'top performance' and 'necessary performance' is becoming the crux of pricing policy.