AI's price crash is 13x a year — only 3x of it is better algorithms

Two studies put a number on the same thing: buying a fixed amount of AI capability is collapsing in price — and almost none of the collapse is the labs getting smarter.

Epoch AI's "The Plunging Price of Thought" finds that reaching a fixed score on AI benchmarks has fallen about 47 percent per quarter since 2023 — roughly 13 times a year — and says no previous transformative technology declined that fast, running 4x faster than DNA sequencing, 6x than compute, 18x than lithium batteries and 54x than electricity.

The worked example is OpenAI's o3. In January 2025 it scored 75 percent on GPQA Diamond, a PhD-level science test, at an estimated 30 cents per question. About 18 months later a GPT-5.6-family model hit the same score for four hundredths of a cent — a 725-fold drop. Epoch's own translation: if cars did that, a $50,000 vehicle would cost $69.

Epoch is careful about what that measures. It is a market price for a fixed benchmark score, set by hardware costs and competition as much as by research, and it rests on five benchmarks spanning math, science and logic puzzles. The group's own words: "reasonable but rough measurements based on the best available data." The decline is also uneven. Game-based puzzles got 7–10 times cheaper a year, math 16–19 times cheaper — and the one benchmark Epoch keeps secret to stop labs training for it fell slowest, which fits the test-gaming critique without proving it.

Then comes the de-confounding. MIT researchers, using pricing data from Artificial Analysis across April 2024 to November 2025, see costs falling 5x to 10x a year. Strip out cheaper hardware and competitive pressure — the latter by looking at open-weight models separately — and the pure algorithmic gain is about 3x a year, close to what earlier efficiency studies found. Their sharper finding runs the other way: running today's frontier model often costs more per correct answer, climbing 3x to 18x a year as reasoning models burn more compute per task. On GPQA Diamond, MIT estimates roughly half of measured progress is associated with rising inference prices.

The take: "AI is getting cheaper" is two claims sharing one sentence. Matching last year's top capability is nearly free now — that is the fact that makes agent deployments pencil out at all. Running this year's best model is getting more expensive per answer, and the price of a benchmark score is not the price of a good decision. We've been tracking the other side of this ledger all month — Gartner lifts 2026 AI spending forecast to $2.7 trillion.

ElevenLabs says it is pacing at $600 million in annual recurring revenue and is reportedly valued by its backers at $22 billion — about double its $11 billion round in February — and its CEO thinks businesses should tell you when the voice on the line is not human.

Mati Staniszewski told TechCrunch that 55 percent-plus of revenue is classic enterprise, that Klarna runs first-line phone support for 35 million US customers on it, and that customers include Deutsche Telekom, Cisco, Adobe and a growing list of governments. On disclosure he is on the record: "I think there should be disclosure at this time" — with the caveat that in five years, when everyone has an agent calling on their behalf, the norm flips. He declined to detail gross margins, saying he does not mind them compressing if it buys market share, and laughed off a reported 2028 IPO timeline rather than confirm it. Hold the valuation loosely: the $22 billion figure traces to July reporting on a tender offer, the terms were preliminary, and no confirmation has surfaced that the sale closed.

US and Canadian regulators recalled the Inmo Air3 smart glasses because the left temple can overheat during extended use — and the remedy is a free firmware update, not a refund or a repair.

The Consumer Product Safety Commission counts 1,643 units sold in the US plus 120 in Canada between December 2025 and August 2026 for between $900 and $1,300, and 10 reports of overheating that include a burning sensation on the face and left ear. Health Canada notes no injuries were reported. Firmware version 3.16 is meant to run the glasses cooler. That a burn hazard is answered over the network says something about the class of device AI glasses are: sensors, batteries and a radio packed into a frame, sold partly on what the software can do and serviceable the same way. The heat was not a surprise to buyers — the company had already shipped a protective sleeve, and review complaints go back about seven months, per The Verge.

What to watch: whether frontier per-answer cost bends, or whether the buying decision stays "cheapest model that is right often enough" while matched-capability prices keep halving.

Would you accept a firmware patch as the fix for a device that got hot enough to burn your ear? Tell us in the comments.

Sources: Epoch AI — The Plunging Price of Thought · MIT FutureTech — The Price of Progress (arXiv) · The Decoder · TechCrunch · Bloomberg · ChosunBiz · US CPSC recall · Health Canada recall · The Verge