AI Performance Costs Falling 13x Annually, Reshaping Generative Creative Economics
The Development
Epoch AI and MIT researchers have published separate analyses confirming that the cost of reaching a fixed AI performance benchmark is dropping at an extraordinary rate. Epoch puts the decline at approximately 47 percent per quarter — around 13x per year — faster than any comparable transformative technology on record. Using OpenAI's o3 as a reference point: in early 2025, hitting 75 percent accuracy on the GPQA Diamond PhD-level science benchmark cost roughly 30 cents per question. By mid-2026, a GPT-5.6 family model reached the same score for four-hundredths of a cent — 1/725th of the original price. MIT's analysis, drawing on Artificial Analysis pricing data from April 2024 through November 2025, is more conservative, estimating a 5x–10x annual decline. When hardware and competitive-pricing effects are removed, MIT isolates algorithmic efficiency gains at approximately 3x per year. The gap between the two figures reflects methodology, not contradiction.
Our Take
The commercial implication is more significant than the benchmark headline suggests. Generative creative at scale was previously gated by cost — teams rationed AI generation to high-ROI use cases because per-query economics forced prioritisation. That constraint is dissolving. The practical effect is that any creative workflow still organised around scarcity assumptions — limited variations, selective personalisation, periodic refresh cycles — is now operating on an outdated model. The MIT finding that competition drives more of the price decline than algorithmic progress is also telling: this is partly a market structure story, and the pressure intensifies as more capable open models keep arriving. Costs will not stabilise.
What Changed
Marketing and creative teams now have access to last year's frontier-level AI output at prices that make volume-constrained workflows obsolete. The capability isn't new — the cost structure is. Output that required selective, high-value deployment twelve months ago can now be run at industrial scale across entire content pipelines.
Marketing Impact
Creative operations and content production are the immediate pressure points. The economics of dynamic creative optimisation, personalised asset generation, and multilingual content scaling shift materially when per-query costs drop by orders of magnitude. Teams still running quarterly creative refresh cycles have a structural cost disadvantage versus those running continuous AI-generation pipelines.
Competitive Implication
In-house creative teams with established generative pipelines gain compounding advantage as costs fall — their fixed investment in tooling and workflow now yields more output at lower marginal cost. Agencies pricing AI-assisted creative on a per-asset or day-rate model face direct margin compression as clients internalise that the underlying generation cost has collapsed.
Strategic Outlook
OpenAI's GPT-6 Sol and Luna models, launched days ago, have likely pushed costs lower still. As open-weight models close the performance gap with frontier proprietary models, the competitive pricing pressure MIT identified will accelerate. Expect per-query costs to continue halving on a sub-annual basis through 2027, making today's cost assumptions obsolete within two to three planning cycles.
The Exploit
Action Item
Creative operations leaders should run a cost audit of their current AI generation stack against September 2026 pricing — any contract or vendor arrangement benchmarked against 2025 rates is overpaying, and renegotiation or model substitution is warranted now, not at renewal.