Skip to content
CostPerPrompt

AI API Price Trends: What the Tracker Actually Recorded

Every article about AI pricing says "prices keep falling." Few can show you receipts. We snapshot the price of every model we track (302 of them) every single day, and this page is computed directly from those snapshots — every number below is a recorded change, not a vibe. Tracking since Aug 2, last updated 2026-08-23.

The headline numbers

The biggest cuts on record

ModelDateOutput /1M beforeOutput /1M afterChange
Alibaba (Qwen) Qwen3 14B Aug 10 $0.91 $0.24 -74%
DeepSeek DeepSeek V4 Pro 0423 Aug 22 $3.20 $0.8951 -72%
Z.ai (GLM) GLM 5.2 Aug 9 $0.792 $0.22 -72%
OpenAI GPT-5.6 Sol Pro Aug 21 $30.00 $10.00 -67%
OpenAI GPT-5.6 Sol Aug 21 $30.00 $10.00 -67%
OpenAI GPT-5.6 Sol Pro (batch) Aug 21 $15.00 $5.00 -67%

Every model page on this site now carries its own price history section (for example, the biggest mover above: Qwen3 14B price history) — check it before hard-coding any rate into a budget.

Why prices move this fast

What to do with this information

  1. Re-shop monthly. The cheapest-APIs ranking recomputes daily; a routing decision older than a quarter is running on stale assumptions.
  2. Watch your model's page. Each model's history section shows when it last moved — a recent cut on a sibling tier is a signal to re-run your comparison in the API cost calculator.
  3. Budget with a corridor, not a point. Given the median cut above, a conservative budget assumes today's rate while an honest forecast allows for downside in your favor — the opposite of most cloud line items.

Frequently asked questions

How often do AI API prices actually change?

More often than the pricing pages suggest. In our tracking window (since Aug 2), 55 of 302 tracked models — about 18% — repriced at least once. Most moves happen quietly, with no announcement: the listing simply changes one day.

Do AI API prices only go down?

Mostly, but not only. Of the 137 repricings we recorded, 66 were cuts and 60 were increases. The direction skews down hard, but increases do happen — usually on preview models graduating to stable pricing, or capacity-constrained flagships. Never hard-code an assumption that a rate can only improve.

How big is a typical price cut?

The median cut in our window was -21% on the output rate — and the tail is dramatic: the single biggest was -74% (Alibaba (Qwen) Qwen3 14B). Cuts of this size flip cost-optimization decisions overnight, which is why a routing setup you tuned last quarter is probably mistuned today.

How should my team react to price volatility?

Three habits: keep model choice in config (not code) so a swap is a one-line change; re-run your cost comparison monthly — our calculators always use current rates; and if a model you depend on cuts prices, check its siblings — vendors often reprice a family together, and the tier below may suddenly clear your quality bar at a fraction of the cost.