A price cut does not rescue a dominated model

Qwen3.8 27B's input price fell 49% between 11 and 15 September, from $0.42 to $0.21 per million tokens. It is still the wrong thing to buy, and the cut does not change that. At $0.087, DeepSeek V4 Flash 0423 scores the same on the intelligence index and generates output 4.8 times faster.

That is worth separating from the usual caveat about price cuts. The normal advice is that a cut to input price only pays off if your workload is input-heavy — long prompts, short answers — and otherwise you are celebrating a discount on the axis you barely use. True, but it leaves the decision open: it depends on your traffic. This one does not depend on your traffic. Qwen3.8 27B at $0.21 is beaten on price, on measured quality and on throughput simultaneously, by two models that were sitting there before the cut landed. There is no prompt shape, no context length and no batching strategy that makes a dominated model the right answer. It is just worse.

Prices and listings from OpenRouter; the intelligence index and throughput from Artificial Analysis.

Throughput is the number that binds

The reason domination is easy to miss is that the cut moves the number everyone watches and leaves the number that actually constrains a deployment untouched. Input price governs the bill. Output throughput governs whether the thing finishes.

At 45 tokens per second, a 2,000-token answer from Qwen3.8 27B takes 44 seconds to generate. The same answer from DeepSeek V4 Flash 0423, at 215 tokens per second, takes 9.3 seconds. Put that inside an agent loop of ten turns and you are comparing seven and a half minutes against a minute and a half. If anything downstream has a timeout — a request handler, a queue worker, a human — the cheaper-per-token model is the one that fails, and the 49% saving is $0.10 per million input tokens that you never get to spend.

This is the part the value ranking cannot show you. Index points per dollar is a genuinely useful sort, and Qwen3.8 27B's cut moved it to 18th of 155 measured models on that measure. But the ratio has no time in it. A model that is slow enough to be unusable ranks perfectly well on quality per dollar right up until you try to serve traffic with it, because the denominator is dollars and the numerator is a benchmark score, and neither of them is a clock.

Why the cut still looked like news

Ten models were cut by 1% or more in those four days, five of them Qwen. The deepest was Qwen3 235B A22B Instruct 2507, down 60% from $0.22 to $0.09. Llama 3.1 70B Instruct fell 44% to $0.40. On a price list, that is a tier repricing itself and it looks like the cheap end of the market got substantially better.

Check the measurements and the picture thins out. Of those ten cut models, three appear in the top forty of the 155 models with a measured intelligence index, and exactly one of the three — Qwen3.8 27B — carries a published throughput figure at all. For the other seven you can see the new price and nothing else. That is not evidence they are slow. It is evidence that nobody has checked, which is a different problem and in some ways a worse one, because an unmeasured model reads as a bargain right up until it is in production.

The background rate here is bad. Of 446 models listed on OpenRouter, 34% have an independently measured quality score and 13% — sixty models — have a published time to first token. For the other 386 listings there is no latency figure in public at all. So the default state of a buyer looking at this catalogue is that they can see price, they can sometimes see quality, and they almost never see speed. A price cut lands on the one axis that is always visible, which is exactly why it moves opinion more than it should.

What to do with a cut

The check is short. When a model you are considering gets cheaper, do not ask whether the new price is good. Ask whether anything already in the catalogue beats it on every published axis at once. If something does, the cut is irrelevant — you were not going to route to it before and you should not now, and the discount is a discount on a decision you had already made correctly.

If nothing dominates it, then the workload-shape question becomes the right one: how input-heavy is your traffic, what does the output length distribution look like, and does the saving survive contact with your actual token mix. That is the harder analysis, and it is worth doing. But it is the second question, not the first, and running it on a dominated model is arithmetic in service of a conclusion that was already settled.

The one thing this week's cuts genuinely changed: GLM 5.3 Flash and DeepSeek V4 Flash 0423 did not move, and they are still the two cheapest routes to a measured index above 30. A wave of discounts elsewhere did not touch them, which is its own kind of signal about where the cheap tier actually is.

All figures measured between 11 and 15 September 2026, the full depth of history available. Prices and context windows from OpenRouter; intelligence index, time to first token and throughput from Artificial Analysis. Quality scores are measured on someone else's tasks, not yours — they narrow what to evaluate, they do not settle it.

Both projects are on GitHub and PyPI. Install them.