Inference price at fixed capability (halving time)
Halving time of the lowest price per million tokens among models at least as good as GPT-4 on GPQA Diamond, fitted from 2023. Epoch's price-trend series while Epoch covered new models (it stopped in February 2025), then each later model on Artificial Analysis that clears GPT-4's own score there and undercuts every model before it. Derived metric inference_price_halving_days.
Reason through thisThe reading
Few points, and the cheapest model at a fixed threshold is rarely the frontier; the METR-hour price series on OpenRouter is the companion reading.
1 reading drawn from the derived metric inference_price_halving_days; 19 observations in the evidence.
Derived (inference_price_halving_days) from: Epoch AI LLM inference price trends; Artificial Analysis API (Intelligence Index, blended price, speed). Epoch AI, LLM inference price trends (CC BY 4.0). Artificial Analysis.
The numbers
| as of | reading | value | record |
|---|---|---|---|
| 2026-04-03 | 5.9 12 0.781 | 121 days | reported |
Why it matters
Price per unit of capability on a benchmark axis, the lab layer's capture proxy; the brief cites about 40x a year at this threshold.
What cuts against it
Few points on one benchmark threshold; list prices; a threshold fixed in 2023 measures commodity capability, not frontier capability. Artificial Analysis gives today's price, not the launch price, so its points for models released before 10 Sep 2026 probably sit low and the halving probably reads slower than it was, and a later price cut on an old model can make it a record after the fact; the two sources run GPQA with different harnesses, each held to its own GPT-4, so the direction of that difference is unknown.
Epoch: the price of reaching a fixed capability fell between 9x and 900x a year, depending on the milestone.
“The rate of decline varies dramatically depending on the performance milestone, ranging from 9x to 900x per year.”
Andreessen Horowitz: for a model of equivalent performance, inference cost falls about tenfold a year (LLMflation).
“For an LLM of equivalent performance, the cost is decreasing by 10x every year.”
How it is measured
- derived metric
inference_price_halving_days(formula in the semantic layer) - source Artificial Analysis API (Intelligence Index, blended price, speed) · default tier 1 · Artificial Analysis API terms; attribution required
- source Epoch AI LLM inference price trends · default tier 6 · CC BY 4.0
- proxy price · cadence annual · valve invention to product
Normal = halving no faster than every 230 days (about 3x a year, the hardware pace); fast = every 110 days or less (10x a year or more). Between is emerging.
Applied to metric:inference_price_halving_days.
Latest readings
- 2026-04-03 · 121 daysas of 2026-04-03(19 obs)
Why this status, and its history
Epoch stopped updating its price chart in February 2025, so the series now continues on Artificial Analysis: each model released after Epoch's coverage ended that clears GPT-4's own GPQA Diamond score there and undercuts every model before it. On twelve points from 2023 to April 2026 the lowest price halves every 122 days (CI 89-192), just outside the fast band (110 days or less) and short of normal (230 days or more), so the status stays emerging; the interval does not reach either band. The Epoch-only fit read 68 days.
- 2026-09-25emerging to emergingconf 45 → 45 · Claude (model)
Epoch stopped updating its price chart in February 2025, so the series now continues on Artificial Analysis: each model released after Epoch's coverage ended that clears GPT-4's own GPQA Diamond score there and undercuts every model before it. On twelve points from 2023 to April 2026 the lowest price halves every 122 days (CI 89-192), just outside the fast band (110 days or less) and short of normal (230 days or more), so the status stays emerging; the interval does not reach either band. The Epoch-only fit read 68 days.
Weighed against it: Artificial Analysis's points use current list prices, not launch prices, which probably makes the halving read slower; a later price cut can turn an old model into a record after the fact; the two GPQA harnesses differ in a direction nobody has measured.
- 2026-09-10unmeasured to emergingconf — → 45 · the evaluator (automatic)
Epoch's price-trend series for the GPQA Diamond threshold set by GPT-4-0314: the lowest price per million tokens halves every 68 days (CI 38-317) on six points from 2023 to December 2024, inside the fast band (110 days or less, about 40x a year). Single tier-6 source with six points caps the status at emerging; the series ends in December 2024 and is refreshed when Epoch republishes it. Initial seed.
Why this confidence
Rests on 19 observations from Artificial Analysis API (Intelligence Index, blended price, speed), Epoch AI LLM inference price trends; best evidence tier 1, grade A; current.
Confidence is independent of status: 90–95 multiple strong independent sources; 70–89 good evidence, some ambiguity; 50–69 mixed or hard to operationalise; 0–49 limited or vague evidence.