34 new models in one month, and a 100x gap between the cheapest and the dearest

September was the busiest release month on record: 34 models from 14 labs. Nobody needs to track all of them. What matters is the shape of the price list that came out of it, because it changes how you should wire up the work you have already automated.

The bottom of the market now sits at ten cents per million input tokens and fifty cents per million out. The top sits at ten euros in and fifty out. That is a hundredfold spread for tasks that, in a lot of back-office work, produce output a reviewer cannot tell apart.

The frontier got more expensive for the first time in years

One release broke the pattern everyone had got used to. The most capable new model of the month came in at roughly two and a half times the previous frontier rate, the first increase at that tier in years. Meanwhile the cheap tier kept falling, with one provider cutting its efficiency model by about a third on input.

So the curve is pulling apart at both ends. Hard reasoning costs more than it did. Routine text costs less than it did. A company that sends everything to one model is now paying for the wrong half of that curve.

What to do with this

• ✳Sort your automated tasks into two piles: ones where a mistake costs money, and ones where a human reads the output anyway
• ✳Move the second pile to a cheap model and compare a week of output side by side before you commit
• ✳Keep the expensive model for work that chains several steps without a human in the middle, where errors compound
• ✳Check your bill by task rather than in total, otherwise one heavy workflow hides behind a hundred light ones
• ✳Write down which model each workflow uses, because in a market moving this fast you will want to swap them next quarter

The release churn is noise. The price spread is the signal, and it rewards anyone who bothers to match the model to the job.