Which LLM Prices Actually Moved in 70 Days of Daily Snapshots

Two months ago we published a two-week look at whether LLM API prices are actually falling. Sixteen days was enough to puncture the “prices only go down” story, but not enough to say much else — and we said so at the time.

This is the follow-up. We’ve now been recording once a day for 70 days — 2026-06-29 through 2026-09-06, more than four times the earlier window, and the longer record answers a sharper question: which prices moved, when, and by how much. Where the numbers here differ from the earlier article, these are the ones to use; that piece stands as the two-week record, this one supersedes its conclusions.

(Seventy days, not ninety. We only publish windows we’ve actually recorded — the record started on 2026-06-29 and cannot be back-filled.)

Everything below comes from our own daily snapshots. We excluded -latest-style aliases and :batch variants from the counts, because an alias can silently repoint to a different model and a batch variant just mirrors its parent — both inflate “change” counts without anything actually being repriced. That leaves 361 models with at least two daily readings.

Three quarters of the market never moved at all

271 of 361 models — 75.1% — did not change their input or output price once in 70 days. Only 90 moved.

Narrow it to the 298 models that were present for the entire window and the picture is the same: 72 moved, 226 didn’t. Roughly one model in four sees any repricing over ten weeks; the rest are frozen.

This is worth internalizing before you build any budgeting process around price-watching. The market is not drifting. It is stuck, punctuated by occasional jumps. A price you looked up in July is probably still correct in September — and if it isn’t, it’s probably wrong by a lot rather than a little.

Among the movers, up beat down

Of the 90 models that moved, we can sort by net change from first reading to last:

More models ended the window more expensive than cheaper, and the increases were larger. That is the opposite of the received wisdom, and it has now held across both our two-week and our ten-week windows.

The 26 models in the middle row deserve their own mention: they changed price — sometimes many times — and landed back at the number they started with. A single before-and-after comparison would have recorded “no change” and missed the whole story.

The biggest cuts

The Luna line is the most interesting entry, and not because of the headline −80%. It got cheap in two steps at the end of July — down to $0.60 on 07-31, a tenth of where it started three days earlier — and then on 08-18 it doubled back to $1.20. Anyone who checked the price on August 1st and again on September 1st sees a clean −80%. Anyone who budgeted on the August 1st number was off by 2× within three weeks.

The GPT-5.6 Sol cut is the cleanest genuine reduction in the set: $30 → $15 → $10 over four days in mid-August, and it has held since.

The biggest increases

Read the percentages honestly: most of these start from a very small absolute number, where a few cents produces a spectacular ratio. A model going from $0.10 to $0.55 per million output tokens is still cheap.

But “still cheap” is not the same as “safe to plan around.” A team that sized its budget on Qwen3 235B Thinking at $0.10 in late June was paying 23× that by September. Nothing broke, no migration was required, no email arrived. The number simply changed, twice, in the direction nobody warns you about.

Gemini 3.7 Flash is the counterexample that should worry planners most: a mainstream model from a major lab that doubled its output price in a single day on 08-29, with no accompanying model change.

Some prices don’t have a value so much as a range

A handful of models repriced constantly:

GLM 5.2 changed price 96 times in 70 days — more than once a day on average — and finished 1.2% away from where it started. For models like this, “the price” isn’t really a number you can quote; it’s a distribution. If you’re running production volume on one of them, a single spot check tells you almost nothing about your monthly bill.

DeepSeek V4 Flash 0731 shows the pattern up close: $0.28 → $0.18 → $0.28 → $0.18 → $0.28 → $0.08 → $0.12 → $0.10 → $0.09 → $0.18, all inside August.

Changes arrive in clusters

Repricing is not spread evenly across the calendar. It bunches:

July 16th alone accounts for 28 output-price changes across four different labs. Days like this are what a daily record buys you: a weekly or monthly check would show the aftermath but not the event, and would leave you unable to tell one lab’s decision from a market-wide shift.

Whether your price moves depends almost entirely on whose model it is

This is the finding with the most practical value:

Anthropic did not change a single published price in 70 days. Neither did x.AI, Amazon, or Perplexity. OpenAI moved 12% of its catalogue, Google 17%.

Meanwhile roughly six in ten DeepSeek and Qwen models repriced at least once.

That’s a genuine planning input, not a trivia fact. If price stability matters more to you than headline rate — because you’re writing a budget you have to defend, or quoting a customer a fixed price — the split above is the thing to look at. The cheapest models in the market are also, consistently, the ones most likely to change on you.

What a pricing page cannot tell you

Every number in this article is invisible on a provider’s own site. Pricing pages show today. When a price changes, the previous one is simply gone: no changelog, no history tab, no way to reconstruct it later. You cannot back-fill a price you didn’t write down.

That’s the whole argument for recording daily. Ten weeks in, we can say which labs move, which don’t, which dates the market repriced on, and which “cheap” models are actually a moving target — none of which could have been recovered after the fact.

The bottom line

• 75% of models didn’t move at all in 70 days. Stability is the default.
• Among movers, increases outnumbered decreases 39 to 25, and were larger (+42% vs −20% median).
• 26 models moved and ended flat — a before/after comparison would have missed them entirely.
• Volatility is concentrated: one model repriced 96 times; a handful of dates carry most of the change.
• Provider is the best predictor. Anthropic: zero changes. DeepSeek: 64% of models changed.
• Cheap and stable are different products. Pick knowingly.

Each model on computetrail carries its own daily price line, so you can check the specific one you depend on rather than trusting a market-wide story — including this one.

Based on computetrail’s daily snapshots from 2026-06-29 to 2026-09-06 (70 consecutive days). Counts exclude -latest aliases and :batch variants; see our methodology. Large percentage moves on low-priced models often reflect repricing or normalization rather than a market trend.