2026 Q3 LLM API Price Change Report: Who Actually Cut Prices (Jul 1 – Sep 30)
We compared two-week median blended prices from early July and late September across 273 models. Nearly 80% barely moved, ten cuts and ten increases held up, and about 42% of all logged price changes were just aggregator flapping.
1. Why this report, and how we measured
Every quarter brings a fresh round of "provider X slashes prices" headlines. We keep a daily snapshot of every model on our tracker, so for once we can check how many of those cuts are real. The obvious method, counting price changes, does not work, and section 5 shows why: between July 1 and September 30 our data logged 1,488 price-change events, and 42% of them were one leg of a price jumping away and then returning to exactly where it had been. Raw change counts measure noise.
So we compared two quiet fortnights instead. For every model we took the median input and output
price across the daily snapshots of July 1–14 and again across September 17–30, then collapsed the two into a
blended price assuming three input tokens for every output token: (3 × input + output) / 4. We
kept only models priced in both windows. That left 293 models; we dropped 9 ~…-latest alias entries (an
alias just re-points to whichever model is newest, so its "change" is a model swap, not a price move) and 11
zero-priced free or media entries, leaving 273 comparable models.
A median shift alone is still not proof, so each mover had to pass a stability test: the range of its daily blended price in the September window must not overlap the range in the July window, and the model may show at most one A→B→A reversal across the whole quarter. The 3:1 ratio is an assumption typical of chat and retrieval workloads. It matters: Qwen3 Coder 480B raised its input price and cut its output price, so it is down 22.8% at 3:1 and exactly flat at 10:1. Change the ratio in the calculator to see your own number.
The headline result is unglamorous. Of the 273 comparable models, 216 (79%) moved by less than 1%, the median change was 0.0%, and 195 models logged no price-change event at all during the quarter. Ten models passed the stability test with a blended cut of at least 10%; ten passed with an increase of at least 10%. Cuts and increases roughly cancelled.
2. The biggest real cuts
Prices below are USD per 1M tokens, shown as July median → September median. Pro variants of the GPT-5.6 models carry identical prices to their base models in our data, so each row covers both.
| Model | Input $/M | Output $/M | Blended $/M | Change |
|---|---|---|---|---|
| GPT-5.6 Luna (+ Pro) | $1.00 → $0.20 | $6.00 → $1.20 | $2.25 → $0.45 | −80.0% |
| GPT-5.6 Sol (+ Pro) | $5.00 → $2.00 | $30.00 → $10.00 | $11.25 → $4.00 | −64.4% |
| Mancer Weaver (alpha) | $0.75 → $0.40 | $1.00 → $0.75 | $0.8125 → $0.4875 | −40.0% |
| Qwen3 Coder 480B A35B | $0.22 → $0.30 | $1.80 → $1.00 | $0.615 → $0.475 | −22.8% |
| GPT-5.6 Terra (+ Pro) | $2.50 → $2.00 | $15.00 → $12.00 | $5.625 → $4.50 | −20.0% |
| ReMM SLERP 13B | $0.45 → $0.35 | $0.65 → $0.65 | $0.50 → $0.425 | −15.0% |
| Magnum v4 72B | $3.00 → $2.50 | $5.00 → $5.00 | $3.50 → $3.125 | −10.7% |
Six of the ten rows are one OpenAI family, and the median comparison understates how messy its path was. The GPT-5.6 models first appear in our data on July 9, so their July window holds only six snapshots. Sol launched at $5.00/$30.00, the same as GPT-5.5, then dropped to $2.50/$15.00 on August 17 and $2.00/$10.00 on August 22. Luna and Terra fell faster, to $0.10/$0.60 and $1.00/$6.00 by July 30, and then both doubled on August 17 (to $0.20/$1.20 and $2.00/$12.00) in the same repricing that halved Sol. The September medians are the lasting prices; the July 30 lows were not.
Qwen3 Coder is the instructive case. Its input price rose 36% ($0.22 to $0.30) while its output price fell 44% ($1.80 to $1.00), both on July 15. If your workload is output-heavy, that is a cut; if you mostly feed it large repositories, it is a wash or a small increase. The other three cuts (Weaver, ReMM, Magnum) are niche community models. No mainstream open-weight model (DeepSeek, Kimi, GLM) shows a cut that survives the stability test. Their apparent cuts appear in section 5.
3. The price increases
Increases were more numerous than cuts. Before the stability test, 34 of the 273 models rose by 5% or more, against 16 that fell by 5% or more. These ten rows pass the test.
| Model | Input $/M | Output $/M | Blended $/M | Change |
|---|---|---|---|---|
| Tencent Hy3 preview | $0.063 → $0.18 | $0.21 → $0.60 | $0.0998 → $0.285 | +185.7% |
| Llama 3.1 8B Instruct | $0.02 → $0.05 | $0.03 → $0.08 | $0.0225 → $0.0575 | +155.6% |
| Qwen2.5 7B Instruct | $0.04 → $0.10 | $0.10 → $0.20 | $0.055 → $0.125 | +127.3% |
| MiniMax M2.5 | $0.12 → $0.27 | $0.48 → $1.08 | $0.21 → $0.4725 | +125.0% |
| Qwen3 30B A3B Thinking 2507 (and four siblings) | $0.13 → $0.20 | $1.56 → $2.40 | $0.4875 → $0.75 | +53.8% |
| Xiaomi MiMo-V2.5 | $0.105 → $0.14 | $0.28 → $0.28 | $0.149 → $0.175 | +17.6% |
The Qwen3 row stands for five models whose blended price rose by exactly 53.8%: the 30B above plus Qwen3 235B Thinking 2507, Qwen3 VL 30B Thinking, Qwen3 Next 80B Thinking and Qwen3 VL 8B Thinking. The 30B has a single endpoint (Alibaba, $0.20/$2.40 when we checked on October 5), so this is not routing noise. Its old price of $0.13/$1.56 was exactly 65% of today's. That is consistent with a 35% discount ending, though we cannot see the cause.
The small, old, cheap models also got dearer: Llama 3.1 8B on July 15, Qwen2.5 7B on July 30, Hy3 preview on August 15, and MiniMax M2.5 in five steps between July 9 and August 22. MiMo-V2.5's input price rose 33% on July 15 with output unchanged, which is the $0.14/$0.28 pair quoted in our open-weight value piece (our price-war analysis has the MiMo background). If a pipeline depends on a model that was cheap because nobody was watching it, re-check the price.
4. Flagships that did not move
Eleven widely used models had zero price-change events in the whole quarter. Their July and September prices are identical:
| Model | Input $/M | Output $/M | Blended $/M |
|---|---|---|---|
| GPT-5.5 | $5.00 | $30.00 | $11.25 |
| Claude Opus 4.8 | $5.00 | $25.00 | $10.00 |
| GPT-5.4 | $2.50 | $15.00 | $5.625 |
| Gemini 3.1 Pro Preview | $2.00 | $12.00 | $4.50 |
| Claude Sonnet 5 | $2.00 | $10.00 | $4.00 |
| Gemini 3.5 Flash | $1.50 | $9.00 | $3.375 |
| Claude Haiku 4.5 | $1.00 | $5.00 | $2.00 |
| Grok 4.3 | $1.25 | $2.50 | $1.5625 |
| Qwen3.7 Plus | $0.32 | $1.28 | $0.56 |
| MiMo-V2.5-Pro | $0.435 | $0.87 | $0.544 |
| MiniMax M3 | $0.30 | $1.20 | $0.525 |
The pattern behind the table: price cuts arrive with the next model, not on the old one. GPT-5.5 stayed at
$5.00/$30.00 all quarter while GPT-5.6 Sol, first seen July 9 at the identical price, fell to $2.00/$10.00. The
Gemini Flash line shows it most clearly. Gemini 3.5 Flash never moved from $1.50/$9.00. Gemini 3.6 Flash first
appears on July 21 at $1.50/$7.50 and reads $0.75/$3.75 from August 13; 3.7 Flash appears on August 13 at
$0.375/$1.875 and is at $0.75/$3.75 from August 28; 3.8 Flash appears on September 2 at $0.75/$3.75. Moving from
3.5 Flash to 3.8 Flash cuts the blended price from $3.375 to $1.50, or 55.6%, and that is also what the
~google-gemini-flash-latest alias shows. Anthropic's pattern is flatter:
Claude Opus 5 first appears on July 24 at
$5.00/$25.00, the same as Opus 4.8, while Opus 5.5
appears on September 22 at $4.00/$20.00, a 20% lower listed price than Opus 5.
Sonnet 5.5 (September 28) matches Sonnet 5 at
$2.00/$10.00. "First appears" means first snapshot in our data, which can trail an official launch by hours.
5. The aggregator flapping effect
Our prices come from OpenRouter, where a model's listed price is the price of whichever provider endpoint is currently the default. When endpoints join, leave or reorder, the listed price jumps, and it often jumps back. Counting input and output moves separately, across the 556 models with price series in Q3, 134 logged at least one change, and 398 times a price moved and then returned to exactly its previous value (an A→B→A reversal) on 75 models. 620 of the 1,488 events, 42%, are the outbound or return leg of such a reversal. 116 models logged at least one downward move, yet only ten survive the cut test above.
GLM 5.2 is the clearest case. Our October 5 snapshot lists it at $0.024 input and $16.00 output per 1M tokens. OpenRouter's endpoint list that day shows 32 endpoints with input from $0.0224 to $2.31 and output from $1.80 to $16.00. The cheap-input default (Relace: $0.0224/$16.00) is a blended $4.02 per 1M tokens at 3:1. DeepInfra at $0.5625/$1.80 is a blended $0.87, and Z.AI's own endpoint at $1.40/$4.40 is $2.15. The endpoint with 25 times cheaper input than DeepInfra is 4.6 times more expensive at a realistic input-to-output ratio. GLM 5.2 logged 111 change events this quarter. Its median blended price fell 22.6% between our two windows, but its daily blended price varied 2.5 times within the September window alone, so that is not a cut you could have counted on buying.
DeepSeek V4 Pro 0423 shows the same trap in the other direction. Its July median was $0.435/$0.87 and its September median $0.83/$1.66, an apparent +90.6%. But its October 5 snapshot reads $0.2088/$0.4176, which is 52% below July. Its endpoints range from $0.21 to $1.91 on input. Sample a different day and you get a different story.
| Model | Median change Jul → Sep | A→B→A reversals in Q3 | Change events in Q3 |
|---|---|---|---|
| DeepSeek V4 Flash 0423 | −42.0% | 9 | 110 |
| GLM 5.2 | −22.6% | 10 | 111 |
| Tencent Hy3 | −42.2% | 6 | 12 |
| DeepSeek V4 Pro 0423 | +90.6% | 4 | 68 |
| Qwen3.5 397B A17B | +42.9% | 14 | 22 |
| Kimi K2.6 | +27.1% | 9 | 38 |
| Qwen3.7 Max | +18.0% | 4 | 6 |
| NVIDIA Nemotron 3 Ultra | +13.5% | 17 | 22 |
These models have the largest apparent moves and fail the stability test. Hy3 shows the problem in miniature: it reached $0.0825/$0.33 on August 26 and went back up to $0.132/$0.528 on October 5, so its "−42%" no longer describes the current price. Qwen3.7 Max flipped between $1.25/$3.75 and $1.475/$4.425 three times in late July and has sat at $1.475/$4.425 since July 29, so its "+18%" is probably real; it fails only on the reversal count. The last model in the table, Nemotron 3 Ultra, had the most reversals of any model in the quarter.
6. What it means for buyers
Read the output price, and know the endpoint. A headline input price of $0.02 tells you very little when the output price is $16. Compare on a blended price at your own ratio using the calculator, and when you call through an aggregator, pin the provider you priced rather than accepting the default. Our guide to platform price differences explains why identical weights carry such different prices.
Expect savings from migration, not from waiting. The cuts this quarter came with new SKUs and with a few repricings of fresh models. Eleven well-known flagships did not budge. If you want the lower price, move to the newer model and test it, rather than waiting for the old one to be discounted. Use compare to line up old and new side by side.
Do not budget on a one-day price. If you track prices yourself, use medians over a week or two, and flag anything that reverts. For fixed costs, our token pricing guide and the value benchmark give the framework, and choosing an LLM covers the quality side.
A quick audit you can run today. List the models you call, note the input-to-output ratio of your real traffic, and price each one at that ratio on its specific endpoint. Then treat a price move as real only if it persists for two weeks and is not followed by a return. By that rule, this quarter's genuine savings were concentrated in a handful of new SKUs, while the biggest-looking headlines on DeepSeek, Kimi and GLM would not have survived the test.
7. Data notes
Data as of 2026-10-05. Prices are OpenRouter's listed model-level prices in USD per 1M tokens, recorded about once
a day; dates in this report are snapshot dates (UTC). They are routing prices, not official list prices. Some
limits to keep in mind: the two-window comparison cannot see what happens in between, such as the July 30 lows of
Luna and Terra. Models first listed after July 14 (including Claude Opus 5.5, Sonnet 5.5 and GLM 5.3) are not in the
comparison. The six GPT-5.6 variants have six July snapshots, not fourteen. A median can also be fooled by a long
excursion: gpt-oss-120b sat at $0.15/$0.60 for 11 of the
14 September-window days before returning to $0.037/$0.17 on September 29, and its July data holds a near-zero
blip of $0.00000003 on July 13, so we left it out of the tables. The Qwen3 explanation (a discount ending) is an
inference from the 65% ratio, not something we observed. All numbers come from queries against our own
pricing_history table, and our China-vs-US comparison
uses the same source.