Per MillionThe record of what intelligence costs

The Meter That Barely Moved

2026-08-27draft — publish-gated on ethics clearance

Draft 1 — 2026-08-27. Companion to “The Great AI Price Collapse Is Over.” Publish-gated on ethics clearance. Numbers trace to dataset/price_changes.csv via scripts/analysis_output_ratio.py; the audit table is dataset/output_ratio_series.csv.

Input vs output pricing

Every AI price headline you’ve read was about input tokens. “GPT-5 at $1.25 a million!” — that’s the input price. It’s the number providers cut, the number press releases lead with, the number that fell 97.9% in four years and became shorthand for intelligence is getting free.

Your bill runs on a different meter. And that one has been strangely, persistently sticky.

Two lines, one widening gap

In 2021, davinci charged $60 per million tokens in both directions — prompt and completion, one price. Then the two meters came apart:

In Out Ratio
davinci (2021) $60 $60 1:1
GPT-4 (2023) $30 $60 2:1
GPT-4 Turbo (2023) $10 $30 3:1
GPT-4o (2024) $5 $15 3:1
GPT-4.1 (2025) $2 $8 4:1
GPT-5 (2025) $1.25 $10 8:1
GPT-5.5 (2026) $5 $30 6:1
GPT-5.6 (Aug 2026) $5 $20 4:1

While the advertised meter collapsed 97.9%, the output meter’s best price ever was GPT-4.1’s $8 — a fall of 86.7%, and it didn’t stay there. Compare three years honestly: GPT-4’s output cost $60 in March 2023; GPT-5.5’s cost $30 in April 2026. The tokens the model produces got exactly 2× cheaper in three years — over a period when the input meter got 6× cheaper and the industry narrative said costs were in free fall.

Why the sticky meter is the one that matters now

The ratio didn’t drift to 8:1 by accident. It widened precisely as the industry pivoted to reasoning models — models that think in output tokens. A chat reply is a few hundred output tokens; a reasoning trace can run tens of thousands, billed at the output rate, mostly invisible to the user. (Consumption figures are illustrative context, not ledger data — the ledger speaks to prices, and the prices alone make the case.)

Look at what providers did with their reasoning lines: o1-preview launched in late 2024 at $60 output — GPT-4’s 2023 price, reintroduced under a new meter, eighteen months of “deflation” later. Anthropic never even pretended: Claude Opus has sat at $15 in / $75 out — a 5:1 ratio — for two and a half years across three model generations, the stickiest price in the entire ledger.

So the honest accounting of the AI cost story is: the meter everyone watches was cut to nearly nothing while usage shifted onto the meter that was quietly repriced upward — first by ratio (1:1 → 8:1), then, since mid-2025, in absolute terms too ($8 → $30 at the flagship, before August’s trim to $20).

What to do with this

If you buy tokens: evaluate models on blended cost for your mix, and assume your mix is more output-heavy than you think if you’re using reasoning modes. An “8× cheaper” flagship with an 8:1 ratio can bill like its predecessor. The input price is marketing; the ratio is the business model.

If you watch this market: the ratio is now a leading indicator. A widening ratio is a provider shifting margin onto reasoning workloads; a narrowing one (GPT-5.6 just went 6:1 → 4:1) is competitive pressure reaching the meter that matters. We’ll track both meters — and the ratio — monthly.


Method: same flagship lineage as analysis #1, both price fields, every value validated against the price-change ledger before the chart renders. Anthropic/o-series figures from the same ledger (first-listing rows). Rerun: python3 scripts/analysis_output_ratio.py.