“Model prices only go down” is one of those things everyone repeats and almost nobody checks. It’s a comforting story for anyone building on these APIs: whatever you pay today, you’ll pay less tomorrow. We started recording LLM API prices once a day specifically to test that story with receipts instead of vibes. Here’s what the first two weeks — 2026-06-29 to 2026-07-14, across 324 models with at least two daily readings — actually show.
Finding 1: Most prices don’t move at all
Over this 16-day window, only 23 of 324 models — about 7% — changed their input or output price. The other 93% sat perfectly still.
That’s the first myth punctured. On any given fortnight, the overwhelming majority of the market is flat. Prices aren’t drifting downward in a smooth curve; they’re stair-stepped and sticky, holding for weeks and then jumping when a lab decides to reprice. If you’re waiting for gradual relief on a model you already use, the data says you’ll mostly wait.
This also has a practical edge: because prices are sticky, a snapshot is usually “good enough” for a few weeks — but the moment a model does move, it can move a lot, and you won’t get a warning.
Finding 2: When prices move, they move both directions
Here’s the part that breaks the “only down” story. Of the 23 models that changed, 14 went up.
The genuinely clean cuts — same model, same code, just a lower number — were led by Qwen:
| Model | Change | |
|---|---|---|
| Qwen2.5 VL 72B Instruct | input $0.80 → $0.25, output $1.00 → $0.75 | −69% in |
| DeepSeek V3.2 | input & output | −6% |
The most eye-catching “drop,” though, wasn’t a price cut at all — and it’s the more useful story. Claude Sonnet Latest shows output falling from $15 to $10 on July 1. But that’s the day Claude Sonnet 5 launched, and Anthropic’s -latest alias silently repointed from the previous Sonnet (Sonnet 4.6, $3 / $15) to the new Sonnet 5 ($2 / $10). If you were pinned to -latest, you woke up running a different model that happened to be cheaper — no code change, and no warning that the model underneath you had changed.
That’s a second blind spot daily tracking catches and a pricing page can’t: an alias can swap the model beneath you. The price line is just the visible symptom; the real event was a generation change you never opted into.
But the outright price increases were bigger and more numerous:
| Model | Change (output) | |
|---|---|---|
| Qwen3 235B A22B Thinking 2507 | $0.10 → $1.50 | +1,395% |
| Qwen3 235B A22B Instruct 2507 | $0.10 → $0.55 | +450% |
| MiniMax M2.5 | $0.48 → $0.90 | +88% |
| Llama 4 Maverick | $0.60 → $0.80 | +33% |
| DeepSeek V3.1 | $0.79 → $0.95 | +20% |
What the increases really mean (a caveat)
A +1,395% jump isn’t “inflation” — read it honestly. Those enormous percentages land on models that were priced near-free to begin with, and the moves usually reflect repricing and normalization: an introductory or under-priced listing being corrected to a sustainable rate, a provider changing which backend serves a model, or an aggregator catching up to a price that changed earlier. Small absolute numbers produce huge percentages.
But that is the point. The narrative says prices fall; the data says prices get corrected, and correction runs both ways. Teams that anchored their budget to an introductory $0.10 rate on Qwen3 235B Thinking would have woken up to a bill 15× larger. “It’ll only get cheaper” is not a safe planning assumption for any single model.
Finding 3: You cannot reconstruct this after the fact
Every one of these moves is invisible on a normal pricing page. A provider’s site shows today’s number; an aggregator shows today’s number. Once a price changes, the old one is gone — there’s no “price history” tab, and you can’t back-fill data you didn’t record. The only way to know that Claude’s -latest alias used to point to a $15 model and now points to a $10 one, or that a model quietly quadrupled overnight, is to have written it down every day before you knew which models would matter.
That’s the whole reason this site exists. Historical price data can’t be created retroactively; it can only be accumulated. Two weeks in, we can already see the shape of the market — sticky, occasionally violent, bidirectional — and every additional day sharpens it.
The honest bottom line
Sixteen days is a start, not a verdict, and we’ll revisit this as the record grows. But even this early the popular story is clearly too simple:
- Prices are mostly flat — ~93% of models unchanged in a two-week window.
- When they move, up is as common as down — 14 of 23 moves were increases.
- The down-moves are worth catching — Qwen2.5 VL 72B quietly shed 69% of its input price; on a model you already run, that’s free money if you notice.
- The up-moves are worth fearing — introductory pricing gets corrected, sometimes brutally.
- Watch aliases, not just prices — a
-latesttag can swap the model under you (and its price) with no warning. - None of it is recoverable after the fact — which is why we record daily.
So: are LLM API prices falling? Some are. Some are rising faster. Most aren’t doing anything this week. The useful move isn’t to assume a direction — it’s to watch the specific models you depend on. Each one on computetrail carries its own daily price line, so you can see which way yours is actually going.
Based on computetrail’s daily snapshots from 2026-06-29 to 2026-07-14. We record LLM API prices once a day and keep the full history; see our methodology. Percentages on low-priced models can reflect repricing rather than a market-wide trend.