Epoch AI shows AI cost falling 47% per quarter since 2023

A data point released by research organization Epoch AI on September 22, 2026 gained traction among AI researchers this week: for the same level of performance, the cost of running artificial intelligence models has been falling by around 47% per quarter since 2023. The figure was highlighted on X by French researcher Aymeric Roucher, known for his work on AI agents, who recommended the account @EpochAIResearch as a reference for anyone who wants to "clearly understand what's happening" in the sector.

The central point of the Epoch AI post is comparative: this speed of price decline has no precedent among technologies that humanity has classified as transformative. According to the survey, AI is getting cheaper 4 times faster than DNA sequencing, 6 times faster than computing hardware in general, and 18 times faster than lithium batteries. The post also cites a fourth comparison, with data predating 1973, pointing to a difference of up to 54 times, but the full text of that part was cut off by X itself (the tweet's "Show more") and it isn't possible to confirm which technology that reference is.

Who is Epoch AI, and why the data matters

Epoch AI is a nonprofit organization dedicated to tracking and projecting trends in the capability, cost, and scale of AI models, based in California and linked to the effective altruism movement. This matters because it's exactly the point raised in replies to the tweet: one user noted that Epoch "is founded, funded, and operated by effective altruists, who certainly have their own agenda in AI," suggesting that the way the data is framed might favor an optimistic reading of AI's trajectory. It isn't a specific methodological criticism of the 47%-per-quarter figure, but it's a legitimate warning: the source isn't neutral, and it's worth reading the data with that context in mind, even though Epoch AI's historical cost-per-performance series is widely cited by other researchers in the field.

Where this price drop comes from

Epoch AI's data measures cost per unit of performance, not a model's raw price. In other words, it's not that OpenAI, Anthropic, or Google are simply lowering their list prices: it's that, to deliver the same quality of response a model delivered in 2023, you now spend a fraction of the computational cost. This curve combines at least three factors that any engineering team working with LLMs has already felt in practice: smaller, more efficient models (the "mini" and "flash" families from the big providers), quantization and distillation techniques that shrink model size without losing as much quality, and the competitive pressure of a market with multiple players of similar weight fighting for the same enterprise customer.

A concrete example of this dynamic: running a text classification or summary generation task that in 2023 only a frontier model like GPT-4 could deliver with acceptable quality is now possible with models that have a much lower cost per token, whether open weight (such as the Llama, Qwen, or Mistral families) or the cheap tiers of closed providers. That's what Epoch AI is capturing in the metric: the same result, costing less and less to get there.

What changes for those building with AI in Brazil

For product and engineering teams budgeting for LLM API usage, this downward curve has a direct implication: architectures that seemed too expensive for production a year ago (multiple chained calls, agents that run several rounds of reasoning, RAG with heavy re-ranking) become viable within a quarter or two, without the team needing to switch providers or rewrite business logic. This favors those who design systems modularly, swapping the model behind a stable interface, and penalizes those who hardcode calls to a specific model expecting that model's price to remain a fixed parameter of the budget.

The opposite also holds: if cost per token keeps plummeting at this rate, the business case for prompt optimization, expensive fine-tuning, or maintaining in-house inference infrastructure loses strength faster than most teams are planning for. It makes sense to revisit, every few months, architecture decisions that seem solid today only because model X was too expensive for use case Y.

The counterpoint the post itself generated

Replies to Epoch AI's tweet bring a skepticism worth noting. One user responded that AI companies "built their business model assuming they'd be a monopoly" and that, by claiming they're "subsidizing" the cost, they're actually just unable to charge more due to a lack of market power. Another user made a similar point, suggesting that this scenario would indicate that investing in these companies would be "a terrible idea" if margin is being sacrificed to gain position.

It's an angle the cost-per-performance metric doesn't answer: it shows what the developer pays, not whether that price is sustainable for whoever offers the service. The question that remains open, and that Epoch AI itself doesn't resolve in this post, is whether this 47%-per-quarter deflation is a genuine result of hardware and algorithm efficiency gains, or whether part of it is cash burn from providers fighting for market share, something that historically tends to correct itself once competition consolidates.

What to keep watching

Epoch AI's data doesn't come with detailed methodology in the tweet itself, just the chart and the summary, so there's no way to audit here which models and which tasks were included in the comparison. For those who want to go to the primary source, it's worth following the @EpochAIResearch account directly and the organization's website, which usually publishes the raw data behind these charts. For engineering teams, the practical lesson doesn't depend on who's right in the subsidy debate: budgeting AI as a long-term fixed cost, in a market that changes price quarter to quarter, is a risky bet, whether the reason is genuine efficiency or a price war.