WEBVTT

00:00:00.000 --> 00:00:04.460
A cheaper token is not the same thing as cheaper useful
work.

00:00:04.460 --> 00:00:10.850
GPT-four’s original eight-K output price was sixty
dollars per million tokens.

00:00:10.850 --> 00:00:18.730
BEP’s selected basket was twelve dollars and eighty-
three cents on September sixteenth, twenty twenty-six.

00:00:18.730 --> 00:00:20.080
Different models.

00:00:20.080 --> 00:00:21.560
Different dates.

00:00:21.560 --> 00:00:24.450
This is not a quality-matched benchmark.

00:00:24.450 --> 00:00:29.870
It also does not prove that the same task became
seventy-nine percent cheaper.

00:00:29.870 --> 00:00:34.570
The basket uses fixed provider weights and a published
selection rule.

00:00:34.570 --> 00:00:36.150
It is not the whole market.

00:00:36.150 --> 00:00:39.660
For an actual task, count input and output,

00:00:39.660 --> 00:00:40.560
retries,

00:00:40.560 --> 00:00:41.600
tool calls, and

00:00:41.600 --> 00:00:42.820
human review.

00:00:42.820 --> 00:00:45.860
Divide the full cost by accepted results,

00:00:45.860 --> 00:00:49.130
holding quality and latency targets constant.

00:00:49.130 --> 00:00:51.020
Then ask who keeps the savings:

00:00:51.020 --> 00:00:54.730
customers, providers, or users doing more work?

00:00:54.730 --> 00:00:56.950
This chart cannot tell us that.

00:00:56.950 --> 00:01:01.340
That takes repeatable task tests and permissioned
customer bills.

00:01:01.340 --> 00:01:06.316
The dated chart, selections, and source notes are at
BEP Research.