Artificial intelligence isn’t just getting smarter; it’s getting dramatically cheaper, and faster than any other transformative technology geek culture has ever ridden to the mainstream. Over the past three years, the cost of achieving a fixed level of AI performance has fallen around 47% every quarter—roughly a 13× drop per year—according to a new analysis from research lab Epoch AI, a finding amplified in a recent breakdown by Tom’s Hardware. For anyone spinning up NPC dialog, procedural worlds, AI dungeon masters, or virtual production workflows, the price of “artificial thought” is entering freefall.
The Epoch study doesn’t measure raw GPU prices or model training budgets; instead it tracks what it costs to hit specific performance benchmarks, including the Graduate-Level Google-Proof Q&A (GPQA) Diamond benchmark at an 81.25% score or better. Across 222 models and up to 11 benchmarks, the researchers find that the inflation-adjusted cost to reach the same score has plummeted by thousands of times in just a few years, far outstripping the historical declines in electricity, lithium-ion batteries, classical computing, and DNA sequencing. DNA sequencing, long the poster child for tech-driven cost collapse, saw an even larger overall drop—but over about 20 years—while AI is matching that fall in roughly half a decade. Classical compute, the domain of Moore’s Law, fell by hundreds of billions of times between 1940 and 2000, yet the short-term slope of the AI cost curve from 2023 onward is steeper still.
Epoch’s report and accompanying data threads highlight how dramatic this “plunging price of thought” looks in concrete terms at the prompt level. One example cited in the analysis involves OpenAI’s o3 model, which in early 2025 could achieve a breakthrough score on a hard math benchmark for well under a dollar per question, versus newer frontier models hitting the same performance for a fraction of a cent in 2026. The speed of the drop isn’t uniform across tasks: cost falls more slowly for game-style puzzles at around 39–43% per quarter and more rapidly for mathematical problems at about 50–52% per quarter, but the average across these domains still lands near that 47% figure. For brand‑new state‑of‑the‑art performance, Epoch estimates that prices initially fall about 66% per quarter—roughly 75× per year—before slowing to around 32% per quarter two years later as the capability commodifies.
Paradoxically, while the price to run AI at a given skill level is collapsing, the cost to build the bleeding‑edge models behind those benchmarks has been exploding. Epoch’s separate work on training economics suggests that the largest frontier runs have seen their budgets grow by a factor of roughly 2–3× per year since 2016, putting billion‑dollar training runs within reach around 2027 if current trends hold. Data from several major labs indicates that compute—both for R&D and for inference—now makes up more than half of their total expenses, with overall spending sitting two to three times higher than revenue in some cases. A think‑tank analysis covered by Computerworld paints a similar picture: running deployed models may pay for itself, but profits are quickly eaten by the race to fund the next frontier system.
That disconnect between plunging inference costs and soaring training budgets creates a brutal business challenge for labs like OpenAI and Anthropic, both name‑checked in wider coverage of the Epoch findings. If customers know that an AI service will be cheaper and more capable within months, their incentive to lock into one provider diminishes, turning frontier models into fast‑depreciating assets rather than long‑term moats. The report’s question—why stick with today’s chatbot, copilot, or AI DM when the next upgrade will outclass it for less money—goes straight to the heart of subscription‑heavy AI offerings aimed at gamers, streamers, and creative pros. For third‑party platforms reselling access to foundation models, the freefall in “cost of thought” can erode margins unless they layer on proprietary features, communities, or data that can’t be as easily swapped out.
For the broader geek ecosystem, though, the upside is obvious: what was “frontier” AI power in 2023 is rapidly drifting into the realm of hobbyist tools, small studios, and solo creators. That means more accessible AI‑driven NPC behavior, smarter procedural generation, and richer assistive tools for modders, tabletop GMs, and indie animators, all at a fraction of the earlier price. The Epoch team is careful to acknowledge caveats—the AI cost series only has direct data from 2023 onward, with earlier values extrapolated, while historical compute and electricity curves stop at 2001 and 1973 respectively—but even with these limitations, the relative trajectory is stark. For fans, the practical takeaway is that previously unaffordable capabilities, from high‑end language models to sophisticated game agents, are likely to continue dropping into consumer‑level products and open tooling at a pace that makes annual hardware upgrade cycles look glacial.
As always with exponential‑looking charts, it’s risky to assume today’s trend line will run forever, especially as supply chains, regulation, and physics push back on both training and inference scales. Still, Epoch’s data supports a near‑term picture where the bottleneck in AI isn’t the cost of spinning up “thought” but the capital required to invent the next generation of minds—and the imagination to do something compelling with them before the advantage evaporates. For the geek crowd, that’s both a warning and an invitation: the price of raw AI cognition is plunging, but the value will accrue to whoever turns that cheap intelligence into unforgettable games, stories, and tools while the window is open.
