Do LLM API Prices Actually Fall? 59 Archived Pricing Pages Say No

Individual LLM prices almost never change. Six of the sixteen model series we tracked held exactly one price across every monthly snapshot, and 71.6% of all observations sat at their series' modal price. The cost of AI genuinely does fall — the cheapest listed model dropped roughly 6.7x in two years — but every step of that came from a new model launching, not an existing one getting cheaper.

"Inference costs are collapsing" is one of the safest things you can say about AI. It is also hard to check, because every provider silently overwrites its own pricing page. Today's prices are everywhere. Last year's are nowhere.

So we went to the Wayback Machine and reconstructed them. 59 archived monthly snapshots of the pricing pages of OpenAI, Google and Anthropic, from January 2023 to August 2026, parsed into a per-model, per-month price series. The full dataset is available as CSV, and the scraper and parser are described in the method section below.

Six models that never changed price once

We tracked sixteen model series with at least four monthly observations each. Six of them show exactly one input price across every snapshot we hold — not a drift, not a rounding difference, the same number every month:

The ones that look like they moved mostly didn't

The other ten series return more than one value across their window, and the obvious reading is that the price changed. Look at the actual values and a different explanation appears.

GPT-4o returns three distinct prices: $2.50, $3.75 and $5.00. Those are not three prices over time. They are the Batch API rate, a cached-input rate, and the standard rate — all printed on the same page, in adjacent columns. Which one our parser caught depends on how that month's page was laid out. The same pattern explains GPT-4o mini ($0.15, $0.30, $0.60) and Gemini 2.0 Flash ($0.075, $0.10, $0.35).

This is a limitation of the extraction, and we are flagging it rather than dressing it up. But it cuts in a specific direction: it manufactures apparent price changes where none occurred. The true rate of price movement is lower than our raw series suggests, not higher. Where the extraction is unambiguous, the price is flat.

Real cuts happen, and they are large

Two series show a change too big to be tier confusion:

• o3 fell from $10.00 to $2.00 per 1M input tokens between May and August 2025 — an 80% cut.
• GPT-5 fell from $1.25 to $0.25 between September 2025 and March 2026 — also 80%.

So in-place price cuts are not unheard of. They are just rare, and when they arrive they are step changes rather than gradual erosion. Nothing in this dataset resembles a model whose price drifted down 5% a quarter.

Where the falling cost actually comes from

Track the cheapest model on any given month's page and the familiar story reappears. In July 2024 the cheapest listed model was GPT-3.5 Turbo at $0.50 per 1M input tokens. By 2026 it was Gemini Flash at $0.075. That is a 6.7x fall in the price floor in about two years.

Every step of that decline came from a new model appearing at a lower price point, not from an existing model being repriced. The floor drops when Flash ships, not when Pro gets cheaper. Providers hold list prices and compete by launching.

That distinction is invisible if you only look at market-level averages, and it is the whole story if you are the one paying the bill.

Method

• Snapshot discovery: the Wayback CDX API, one 200-status capture per calendar month per provider, from January 2023. Retrieved with the `id_` modifier so the archive returns the original response rather than a rewritten page.
• Providers: OpenAI (29 months), Google (25 months), Anthropic (5 months). 59 of 89 retrieved snapshots yielded usable token prices.
• Unit normalisation: early pages quote per 1,000 tokens and later ones per 1M, and some pages mix both. Units are resolved per price from its surrounding text, not per page, and every value is normalised to USD per 1M tokens.
• Non-token prices excluded: a price only counts if "token" or "MTok" appears near it, which keeps per-image, per-minute and per-request rates out of the series.
• Model association: prices are matched to model names by bounded proximity — each model's search window stops where the next model is mentioned — and a row is only accepted if it carries explicit input and output labels. Unlabelled rows are discarded rather than guessed at.
• Reported value: the modal price per series, not the mean, so a single misparsed month cannot shift a model's headline number.
• Validation: extraction was checked against independently known values, including GPT-4o at $5.00/$15.00, GPT-3.5 Turbo at $0.50/$1.50, o1 at $15/$60, Claude Opus 4 at $15/$75 and Claude Sonnet 4 at $3/$15.

Limitations

Anthropic coverage is thin — five months — and that is a data problem, not an Anthropic one. For most of the window their archived pricing URL captured the consumer subscription page (Pro, Team, Max) rather than API token pricing, so there were no token prices on it to extract. Treat the Anthropic rows as a sample, not a series.

Snapshots are point-in-time. A price that changed and changed back between two monthly captures is invisible to this method. Tier ambiguity remains unresolved for the ten non-constant series, as described above. Mistral and DeepSeek are not included in this cut. And the archive is not a complete record — one OpenAI month returned a 403 and is simply missing.

What to do with this if you are budgeting

The planning assumption that costs will fall under you is wrong at the level you actually operate at. If you build on a model today, that model will very probably cost the same in a year. The savings are real, but they are only available to teams that migrate — and migration has its own cost in evaluation, prompt rework and regression testing.

Which turns a finance question into an architecture one. Systems where the model sits behind a clean interface can capture the floor as it drops; systems with prompts and model quirks baked through them cannot, and end up paying 2023 prices in 2026. That is worth deciding deliberately at the start of an AI automation build rather than discovering it at renewal.