Disclaimer

Vareon ran every test behind the results on this website, analysed the results and wrote the reports. That makes each of them an internal benchmark, including the tests that use a public benchmark.

No result has been independently verified, audited or reproduced by a third party, and none has been certified by a standards body or a testing laboratory.

A figure describes the test it comes from: the work, the model, the settings, what it was compared with and the number of runs. The page for each result sets these out under “How it was run”. See the results

A result from one test does not carry over to another workload, model, dataset or environment. On your own work, the figures may be higher or lower.

Some tests use benchmarks and datasets that others built and published: Τ-bench by Sierra Research, LongMemEval by Wu et al., World of Workflows by Skyfall Research on ServiceNow, and three public intent datasets. Vareon ran these tests itself, under the conditions it reports.

Their authors have not reviewed, verified or endorsed Vareon’s results. Where a result is set against work that others published, such as the frontier models’ predictions on World of Workflows, the comparison is Vareon’s own.

Results describe what happened in the tests. They are not a guarantee, forecast, certification or service-level commitment, and they are not an offer. What a deployment is expected to achieve is agreed in writing for that deployment.

The surest way to know what Duplex does on your work is to measure it there. A pilot runs on your own work and reports its own figures. How a pilot works

GPT, Gemini, Claude, ServiceNow and the other models, products and benchmarks named on this website belong to their owners. A name identifies what was tested or compared, and does not mean that its owner sponsors, endorses or has reviewed Vareon’s results.

Vareon may run a test again, correct a result or withdraw it. The date above shows the latest revision of this page.

For a question about a result or how it was tested, use the Contact page.