More accurate at every setting, at a seventh of the cost.

At all eight of its settings, Duplex was more accurate than a semantic cache at any of its five, or GPT-5.5 alone at its one, on the same 3,000 decisions. Three more runs, on fresh streams, showed the same order.

See it happen →Every result

94.1 to 94.8%Duplex accuracy, at every one of eight settingsThe best a cache reached: 77.3%
84 to 91%On three fresh runs of 1,000 decisionsThe best a cache reached: about 73%

Every setting, on one chart.

GPT-5.5 alone has one setting: 59.8% accurate, at $0.82 per 1,000 decisions.

Duplex at its worst beats a cache at its best.

The least accurate Duplex setting against the most accurate cache setting.

At its least accurate setting, Duplex scored 94.1%. The cache, at its most accurate, scored 77.3%.

Tuning a cache does not buy accuracy.

The same cache, from its loosest matching to its strictest.

Tightening a cache’s matching sends more requests to the model: from 4 calls to 212.

The same order, run after run.

Three fresh runs of 1,000 decisions each, in a second, independent setup.

Over three fresh runs in a second, independent setup, Duplex scored 87.7%, and the best cache 73.0%.

How it was run.

The work: The same customer-operations stream as the cost result: one right action out of about nineteen, or a refusal, with five in-house rule changes.
Settings: Duplex at eight settings, from most willing to answer itself to most cautious; a semantic cache at five, from loosest to strictest matching; GPT-5.5 alone at its only one.
Scale: One 3,000-decision stream, then three independent runs of 1,000 decisions in a second setup.
Kept equal: The same recorded GPT-5.5 answers for every point, the same feedback and the same refusal rules.

What it means for your work.

Steady at every setting.
All eight Duplex settings landed between 94.1% and 94.8% accurate.
Accuracy a cache cannot reach.
Across its settings, the cache topped out at 77.3%, and at 44.2% after a rule change.
Cheaper than the model, everywhere.
Even Duplex’s most cautious setting cost less than GPT-5.5 alone, and was 34 points more accurate.

Decides whether to answer, hand over or decline →