The LLM price war is real at the bottom and imaginary at the top

Across 60 models from 17 providers, input prices span 1000x and output prices 2250x. That is not one market having a price war. It is two markets that happen to share an API shape.

By Soufiane Bellahmar · Founder and editor, LLM Cost Lab
Figures recomputed from the 60-model dataset on every build. Rates last verified 2026-08-16.

"The price of intelligence is collapsing" is one of those claims that is simultaneously true and useless. It is true if you look at the bottom of the market, where a million input tokens now costs less than a cup of coffee costs to make. It is close to false at the top, where the models people actually reach for when the work matters have held their prices with remarkable discipline. Averaging the two together produces a number that describes nobody's bill.

So rather than average, here is the distribution. Every figure below is computed from our own dataset of 60 models, each rate taken from the provider's published pricing page.

The shape of the market

The cheapest input rate we track is Qwen3.7 Flash at $0.03 per million input tokens. The most expensive is GPT-5.5 Pro at $30.00. That is a spread of 1000x between the two ends of a market where every product answers the same HTTP call and returns the same shape of JSON.

The median sits at $1.00. The tenth percentile is $0.10 and the ninetieth is $5.00. The distribution is not a bell curve with a fat middle; it is a dense cluster near zero with a long thin tail. 29 of 60 models — 48% of everything we track — charge under a dollar per million input tokens. Only 2 charge ten dollars or more.

Output tells the same story more violently. The gap between Llama 3.1 8B Instant (via Groq) at $0.08 and GPT-5.5 Pro at $180.00 is 2250x. Output is where the real money goes, and it is where the spread is widest.

Every provider now defends a floor

The most interesting pattern in the data is not any single price. It is that nearly every provider — including the ones whose flagship models are the most expensive things on the list — now ships a deliberately cheap small model. Here is the cheapest option each vendor offers:

Read that as a strategic map rather than a shopping list. A provider whose premium model costs twenty times its own cheap one is not confused about pricing. It is refusing to let a competitor own the entry point, because the entry point is where developers start and switching costs accumulate. The cheap tier is a moat, not a product line.

The practical consequence for you is that staying inside one provider is cheaper than it used to be. A few years ago, cost-optimising a workload usually meant splitting it across vendors. Now the same vendor will usually sell you both the expensive answer and the cheap one, with the same auth, the same SDK, and the same billing relationship. Whether that is a better deal than mixing vendors is a question our model switching calculator will answer for your actual token mix.

The top of the market did not move

Now look at the other end. These are the eight most expensive models by output rate:

Frontier pricing has been notably sticky. The reason is not cartel behaviour; it is that the buyers at the top of the market are not price-sensitive in the way the buyers at the bottom are. If a model is doing legal review, writing production code, or acting as an autonomous agent with real spending authority, the cost of the model being wrong dwarfs the cost of the tokens. A team in that position does not switch to save forty per cent on inference. Providers know this, and price accordingly.

The cheap end competes on price because that is the only axis its buyers care about. The expensive end doesn't, because its buyers are optimising something else entirely.

What this means for a budget

Three things follow from the split, and they are more useful than any headline about prices falling.

1. Your blended rate matters more than any listed price

Almost nobody runs one model. A realistic production system routes easy requests to something cheap and hard ones to something capable, which means your effective cost per million tokens is a weighted average you have probably never calculated. Because the spread is 1000x wide, small changes in that routing ratio move your bill far more than any provider's price cut will. Ten per cent of traffic moving from a flagship to a small model is usually worth more than a year of industry-wide price declines.

2. Falling prices do not reduce bills

This is the counterintuitive one. In practice, every price cut we have tracked has been followed by teams spending the headroom rather than banking it — longer context, more retrieval chunks, more agent steps, reasoning modes left on by default. The unit price falls and the unit count rises to meet it. If you want to actually spend less, the lever is reducing tokens, not waiting for a cheaper rate.

3. The gap is an opportunity, not a problem

A 2250x spread on output means there is almost certainly a model that does your specific job at a fraction of what you are currently paying. Not for every job — but for classification, extraction, routing, summarisation of already-clean text, and the dozens of small internal calls that surround the one call that actually needs to be smart. Those are the calls worth auditing, and they are usually the majority by volume.

If you want the current numbers rather than the argument, the price index carries the full distribution and the comparison table lets you sort all 60 models on any column. To turn this into a figure for your own workload, start with the API cost calculator.