Archive · 2026-09
September 2026,
frozen.
These figures were written once, at month close on 1 Oct 2026, and cannot be edited. Cite them freely: this page will read the same in a year.
What this is: a public monthly record of what AI model prices actually did, which providers moved them, and which model is still the cheapest one good enough for a given job.
Number of the month
+5203%
$0.0040 at OpenInference against $0.210 at Azure, per MTok in.
Direction
6120 recorded price moves in September 2026: 4167 cuts and 1953 rises, so the market moved down on balance. 68% of all moves were cuts.
Four terms this page uses
- MTok
- One million tokens. Every price on this page is US dollars per million tokens, so models with different token accounting stay comparable.
- Price move
- One observed change to a live listed price, recorded in the append-only ledger with its direction. A model appearing for the first time is a new listing, never a move. Full definition.
- Provider spread
- The gap between the cheapest and dearest host serving identical weights. It is a routing choice, not a quality difference.
- Equivalence band
- The score range around your current model, widened by the benchmark's own measurement margin, that a candidate must sit inside before a switch is offered.
Price moves
What changed in September 2026
Recorded from the append-only price ledger: every observed move, per model and per host, on both the input and output side.
- Decreases4167
- Increases1953
- New listings280
- Total moves counts increases plus decreases only. New listings are shown here for context and are never folded into that total.
Top 5 increases
Top 5 decreases
Who reprices most
Ranked over everything we hold. The leaderboard updates as more price moves are recorded.
Market structure
The same weights cost wildly different money
Identical model, different real provider. Aggregator listings are excluded, so every gap below is a genuine provider-to-provider spread you could act on today.
Providers per model, across 412 models with at least one real (non-aggregator) endpoint. Most weights are single-sourced; a small tail is served everywhere, and that tail is where the spread lives.
Quality per dollar
The cheapest model that still clears the band
For each task class we take the leading published score, subtract that evaluation's measured margin, and pick the cheapest model still above the line. Only suites with a real measured margin appear, because a benchmark without one cannot back a claim.
Benchmark saturation
An evaluation stops being usable when the spread between models collapses into the measurement margin. We require the observed spread to exceed twice the margin; a ratio at or below 1.0 means the instrument can no longer tell models apart.
tau_banking · aa:tau_banking
spread 49.69 vs margin ±8.42 across 88 scored models, discriminating.
2.95×vs 1.0× floorterminalbench_v2_1 · aa:terminalbench_v2_1
spread 88.02 vs margin ±10.33 across 88 scored models, discriminating.
4.26×vs 1.0× floorlcr · aa:lcr
spread 86.67 vs margin ±9.46 across 150 scored models, discriminating.
4.58×vs 1.0× floorscicode · aa:scicode
spread 49.50 vs margin ±5.33 across 164 scored models, discriminating.
4.65×vs 1.0× floorgpqa · aa:gpqa
spread 61.20 vs margin ±5.84 across 154 scored models, discriminating.
5.24×vs 1.0× floorhle · aa:hle
spread 59.60 vs margin ±1.67 across 164 scored models, discriminating.
17.89×vs 1.0× floor
Newsletter
These numbers, weekly.
One email a week: the price moves we recorded, which switches cleared quality, and what changed for the cheapest model good enough for a given job. No pitch, unsubscribe in one click.
Method
How a switch gets measured
Price sync at freeze
The prices on this page are the catalog as it stood when this month was frozen. The engine itself keeps re-syncing from public provider feeds; the live page reflects that, this archive deliberately does not.
Independent benchmark scores
Quality comes from published third-party evaluations, per task class. We do not run our own private eval and we are never paid for placement.
The equivalence band
A candidate model only qualifies when its score sits inside the band around your current model for the task class in question. Cheaper-but-worse never clears.
Measurement margin
Every score carries its own uncertainty. We compute the real margin and require the gap to survive it before a switch is offered.
Latency ceilings
Median latency is part of the decision, not an afterthought. Set a ceiling and candidates that breach it are dropped before cost is even compared.
Refusals with reasons
When nothing clears, you get a stated reason, not a weaker suggestion. A quiet downgrade would cost you more than the saving is worth.
Cite and reuse
Take the numbers with you
Every figure here is free to quote, repost or chart, with attribution to CostMyAI. We would rather be the cited source than the hidden one.
Citation
CostMyAI Intelligence, September 2026. Retrieved 1 Oct 2026. https://www.costmyai.com/intelligence/2026-09
Permanent link
https://www.costmyai.com/intelligence/2026-09?ref=share&card=cite
Download the table
Same figures, same identifiers as the anchors on this page, so a chart you build from the file matches the page it came from.
Every figure card on this page carries its own image and a ready post with the source line already in it.
Embed
Put the live market on your own page
A self-contained frame that rotates the three sharpest numbers of the month: how many prices moved, the steepest rise and the steepest cut. It refreshes itself, needs no script on your site, and always credits CostMyAI.
<iframe src="https://costmyai.com/embed/intelligence-widget"
title="AI price market, via CostMyAI"
width="100%" height="200" loading="lazy"
style="border:0;max-width:520px"
referrerpolicy="strict-origin-when-cross-origin"></iframe>- RotationMonth-over-month move count, biggest rise, biggest cut.
- FreshnessServer-cached and refreshed every five minutes.
- SafetyIsolated iframe, no script in your page, nothing configurable.
- AttributionThe “via CostMyAI” link is part of the widget on every plan.
Archive
Every closed month, frozen and permanently linkable
At 00:00 UTC on the first of each month we write that month's final figures once and never touch them again. A correction is filed as a new restatement row that points back at the original, so the number you cited stays exactly as you cited it.