Village
Log in
Two broad bands of fine hatching cross like open scissors: a graphite band falls from the upper left to the lower right, and a plum band rises from the lower left to the upper right.

The ten best AI models cost eight times as much per task as two years ago

The ten most capable AI models cost a median of eight times as much per task as the top ten did two years ago, even as matching last year’s best model now costs 1% of its launch price. At default reasoning settings the rise is fivefold; the highest settings make it eightfold.

by Village Team

Using the most capable models is getting more expensive. Village’s LLM Atlas scores each model’s capability at each reasoning setting and estimates its cost for one task from two agentic benchmarks: Terminal-Bench 2.1, where models fix code in a terminal, and GDPval, where they produce professional documents. By that measure, the median cost of the ten most capable models, which we call the top ten, has risen eightfold in two years.

The top ten’s cost per task rose eightfold in two years

Cost of one agentic task (Terminal-Bench 2.1 and GDPval workloads) on a log scale, monthly from June 2024 to October 2026. The median cost of the ten most capable models, each at its most capable setting, rose from 21¢ in October 2024 to 83¢ in October 2025 and $1.77 in October 2026; a shaded band spans the middle half of the ten. Over the same time, matching October 2024’s best (o1 Preview) fell from $2.63 to 1.9¢, and matching October 2025’s best (GPT-5 Pro (high)) fell from $4.19 to 4.3¢.

  • The top ten now cost a median of $1.77 a task. That is up from 21 cents in October 2024.
  • Matching last year’s best model now costs 1% of its launch price. GPT-5 Pro (high), the most capable model in October 2025, cost $4.19 a task; the cheapest model that scores as high today costs 4.3 cents.

The gap between last year’s best and this year’s top ten is now about 40-fold: 4.3 cents a task against $1.77.

Max reasoning, first offered in February, now accounts for eight of the top ten models

Part of the rise comes from reasoning settings, which control how long a model thinks before it answers. Every token of thinking is billed, and labs keep adding higher settings.

A grid with one column per month from June 2024 to October 2026 and one square for each of the ten most capable models, colored by the reasoning setting at which it scores highest. o1 adds low, medium and high, December 2024. GPT-5.1-Codex-Max adds xhigh, November 2025. Claude Opus 4.6 adds max, February 2026. October 2024: 10 no effort level. October 2025: 4 no effort level, 6 low to high. April 2026: 1 no effort level, 1 low to high, 6 xhigh, 2 max. October 2026: 2 low to high, 8 max.

  • Each new setting reached the top ten within months. Max first appeared in February 2026; today eight of the top ten score highest at it.
  • Each step up costs more than the last. Across models that offer both, max costs a median 60% more per task than xhigh, more than twice what xhigh adds over high.
  • Together, the settings add about two-thirds to the top ten’s cost. Counting only each model’s default setting, that cost rose fivefold in two years rather than eightfold.

Unlike a model’s price, its reasoning setting is the buyer’s choice, and the highest setting is the poorest value.

Moving Claude Opus 5.5 from xhigh to max nearly triples its cost per task for just 0.6 capability points, on a scale where the top ten span about three.

GPT-6.1 Sol at xhigh scores within two points of the top ten’s leader for 41 cents a task, less than a quarter of the top ten’s median cost. For most work, a setting one or two steps below the maximum gives nearly the same capability for a fraction of the cost.

How we built this

This study uses Village’s LLM Atlas, which estimates a capability score for each model and reasoning setting from public benchmark results and, with a separate cost model, the typical dollar cost of a task from evaluators’ runs, described in its methodology. Agentic tasks are the cost model’s Terminal-Bench 2.1 and GDPval workload. On each date, the top ten are the ten most capable models released by then, each at its most capable setting with a cost estimate; their cost is the median of those ten. A level’s cost is the cheapest setting released by then that scores at least that high. The scores and costs are from November 2, 2026; Atlas updates as new results arrive.

Settings without a cost estimate are left out, including premium tiers such as GPT-6 Astra Pro, so the rise is if anything understated. A setting’s arrival is the release of the first model whose lab offered it. GPT-5.1-Codex-Max, the first with xhigh, has no cost estimate, so xhigh reaches the chart’s top ten a month later with GPT-5.2. Agentic costs for 2024 models are the cost model’s estimates rather than measured runs. Using the five most capable settings instead, or only models scored on at least 10 benchmark families, gives a similar rise.

Next up