The Great AI Price Collapse Is Over

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

Draft 1 — 2026-08-27. Publish-gated on ethics clearance. Every number below traces to dataset/price_changes.csv (5,548 provenance-carrying price events, 2021–2026) via scripts/analysis_deflation.py; date confidence levels are in dataset/verification.md.

For five years, the loudest constant in AI was the price collapse. In late 2021, a million tokens into OpenAI’s best model cost $60. By August 2025 — GPT-5’s launch — the same million tokens into the flagship cost $1.25. A 97.9% collapse in 44 and a half months: the price of frontier-grade intelligence halved every 8 months, faster than transistor economics ever managed, and everyone built their plans around the assumption that it would keep doing that.

It stopped. August 2025 was the bottom, and the data says so plainly.

The turn

Trace OpenAI’s flagship lineage — davinci, GPT-4, 4-Turbo, 4o, GPT-5, and the 5.x line — and the input price does this:

From the bottom, the flagship price quadrupled in eight and a half months — a doubling every 4.3 months, faster going up than it ever fell coming down. GPT-5.5 pricing landed back at 2024’s GPT-4o launch price. And the top of the range has fully round-tripped: GPT-5.5-pro lists at $60/MTok — to the dollar, what davinci cost in 2021.

This isn’t one company’s quirk. DeepSeek — the price floor of the whole market, the lab whose $0.14 listing in May 2024 forced everyone’s hand — raised API prices roughly fourfold with the V4 family (announced 2026-07-31, effective 2026-08-16: v4-flash input $0.14 → $0.44, output $0.28 → $1.32 at peak hours), reintroducing the peak/off-peak split it had abolished a year earlier “to allocate resources more reasonably.” When the price floor and the price ceiling turn in the same year, that’s a regime change, not noise.

What this is not saying

Honesty about the frame, because the frame is doing work:

• This is the price of the best available model, not constant quality. GPT-5.6 is not GPT-5, and paying 2024 prices for 2026 capability is still a deal. Constant-quality deflation continues: yesterday’s frontier keeps getting cheaper (GPT-5-class pricing survives in the 5.6 line’s cheaper variants at $1.00–2.50).
• The commodity tier never stopped collapsing. Nano-class models list at $0.05–0.20. What’s actually happening is dispersion: in 2021 OpenAI sold intelligence at one price; today its own lineup spans $0.20 to $60 — a 300× spread. “What does AI cost?” stopped having a single answer, which is precisely why frontier budgets are blowing up while demos stay cheap.
• This is input pricing. Output tokens — where reasoning models actually bill — are a second, uglier story (analysis #2).
• 2026 dates are ledger observation dates (community-file commits, which we verified track announcements to the day for headline launches). DeepSeek events are cited from primary sources, not the ledger: the community file carried wrong V3.1-era values ($0.60 vs the real $0.56) and observed the V3.2 cut 3.5 months late — see dataset/verification.md, “Verified with corrections.”

Why it turned — and what to watch

The candidate explanations aren’t mysterious: the reasoning era made inference genuinely expensive; compute stayed scarce; and the land-grab phase — where API pricing was a customer-acquisition subsidy — ended when the customers arrived. The ledger can’t adjudicate motives. What it can do is what deflation-narrative pieces never do: track this forward, publicly.

Watchables, revisited monthly: whether GPT-5.6’s $5.00 holds through the next flagship; whether DeepSeek’s V4 raise sticks or gets competed back down; whether Anthropic and Google follow the frontier repricing or hold. If the $5.00 flagship price survives to mid-2027, the 8-month halving era is formally history — and every business model priced on “tokens trend to free” has a problem it hasn’t accounted for yet.

Method: flagship lineage and all prices from the price-change ledger (sources: Wayback Machine snapshots 2021–2023, LiteLLM pricing-file git history 2023–2026, provider announcements for date anchoring). The series, with per-point provenance refs: dataset/deflation_series.csv. Rerun: python3 scripts/analysis_deflation.py.