# Epoch AI Puts the Fall in the Cost of a Benchmark Answer at 725-Fold in Under 18 Months

OpenAI's o3 needed about 30 cents per question to reach 75 percent on the graduate-level science benchmark GPQA Diamond. GPT-5.6 Luna matched that score for four hundredths of a cent.

- Published: 2026-09-23T11:26:37.007Z
- Canonical: https://polylog.news/ai/2026-09-23/epoch-ai-puts-the-fall-in-the-cost-of-a-benchmark-answer-at
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Epoch AI](https://epoch.ai/publications/the-plunging-price-of-thought)

The research group Epoch AI has quantified a cost decline that engineers had previously only seen reflected in their invoices. In its analysis [The Plunging Price of Thought](https://epoch.ai/publications/the-plunging-price-of-thought), Epoch estimates that OpenAI's o3 model reached a 75 percent score on GPQA Diamond, a multiple-choice examination of doctorate-level physics, chemistry and biology, at an average cost of about $0.30 per question. Just under 18 months later, GPT-5.6 Luna reached the same score at roughly $0.0004 per question, a 725-fold reduction.

The finding [circulated widely](https://t.me/aipost/8231) among practitioners this week, and it is worth reading carefully. The measure is the cost to reach a fixed score on one benchmark, not the cost to perform a fixed unit of useful work. GPQA Diamond has little room left for scores to increase further, a limitation Epoch itself has written about. A 725-fold drop in the price of one answer is not the same as a 725-fold drop in the price of the reasoning a production system actually performs, where long contexts, tool calls and retries dominate the bill.

What the number does establish is the direction and scale of the decline. It comes from three compounding sources: cheaper serving hardware and better software, smaller models distilled to match the performance of larger ones, and post-training methods that produce correct answers using fewer generated tokens. None of the three shows signs of stopping. The same pattern appeared in this week's model launches, where OpenAI and Anthropic both delivered cost reductions that came largely from token efficiency rather than list-price cuts alone.

## What this means

If capability per dollar keeps falling at anything near this rate, the binding constraint on deploying AI shifts from the inference budget to everything around it: data access, evaluation, integration and legal liability. That favors application companies with strong distribution and proprietary data over companies whose advantage was the ability to afford expensive model calls. It also weakens the argument that frontier capability can be limited by price, since last year's frontier score becomes this year's low-cost commodity.

## What to watch

- Whether Epoch publishes an equivalent cost curve for agentic, multi-step benchmarks rather than single-question exams, which would test whether the same decline holds for the workloads engineers actually run.
- Whether the labs' reported token counts per completed task keep falling, since that part of the cost decline does not depend on subsidized pricing.