Key findings from the AI models analysis
The ranking prices 432 listings against every company that hosts them, across 1,810 host endpoints. The figures below count the 356 distinct models behind those listings, with batch tiers folded into the model they discount. These are the patterns that came out of it.
114 of 123 models cost more than an equally capable option
The median one costs 9.2× what it needs to. The worst, Claude Opus 4.1, costs 351× more than Solar Mini 4 for the same work at the same score.
What 1,000 support replies cost on each. Left is the model you were going to use, right is the cheapest one that scores as well.
Every beaten model, grouped by how far past the cheapest equal it sits.
How this was measured. A model counts as beaten when another scores within 18 Elo of it or higher in direct comparisons and costs less for the same workload. Batch billing tiers fold into the model they discount, so nothing is reported as beaten by itself.
33% of prices depend on who hosts the model
116 of 356 models vary by more than 1.2× across hosts. The widest, GLM 5.3, runs from $0.06 to $2.80 per million input tokens. Identical weights, 50.9× apart.
cheapest hostdearest hostSame weights, same model. Log scale.
How this was measured. Prices read per provider endpoint from OpenRouter. Range is min to max across every host currently serving that model; the marker shows the median.
10% are quietly served at 4 bit precision
37 of 356 models have at least one provider running fp4 or int4 weights, including Kimi K3, the highest scoring model in the catalogue. Compression at this level measurably degrades output, and no major leaderboard reports it.
The highest scoring models with at least one host running compressed weights. The bar is how many companies serve it, which is how many ways your call can land on the compressed one.
How this was measured. Quantization is read from each provider endpoint's declared precision. A model is flagged when any host serves fp4 or int4, even where others serve it at full precision.
15× is how far a context window varies by host
34 models vary by more than 1.5× across hosts. Qwen3.8 27B runs from 66K tokens on one provider to 1000K on another. The figure a model is marketed with is usually the maximum, available from a single host.
smallest windowlargest window offeredSame weights, same model. Log scale.
How this was measured. Context read per provider endpoint. Range is min to max across hosts; the marker shows the value most hosts actually serve.
25% bill reasoning tokens you cannot switch off
89 of 356 models cannot turn reasoning off, and several default to maximum effort. Claude Opus 5.5 and Claude Sonnet 5.5 are both on the list. The published output price is a floor, not the amount you pay.
The highest scoring models that always reason. The figure is the output rate per million tokens, which is what the reasoning is charged at.
How this was measured. Taken from each model's declared reasoning configuration: flagged where reasoning is mandatory rather than optional.
90% is the cached input discount, rarely compared
233 of 356 models publish a separate cache read rate. On any workload that asks repeated questions about one document this term dominates the bill, and it reorders which model is cheapest.
Discount on cached input for the highest scoring models that publish one. The median across all 233 is 90%, which is the number in the heading.
How this was measured. Cache saving is the published cache read rate against the standard input rate for the same model. The headline is the median across the 233 models that publish both.
Every figure here is arithmetic on public data that needs no key or account, and it is redrawn from scratch each time the dataset refreshes. Take the numbers and check them yourself: the full dataset is free to download under CC BY 4.0, and the ranking lets you change the workload and watch the order move.
Updated 12 October 2026, from 356 models across 1,810 host endpoints. Free to quote, reprint or chart with attribution to BestHunt.