Predicting what changes next.
Step 1 of 4: An operation
Beat three frontier models at predicting system changes.
World of Workflows is a public enterprise benchmark built on ServiceNow: after each operation, predict how the system’s records change. Duplex learned the workflows as it went, with no model calls, and was scored by the benchmark’s own code.
World of Workflows, by Skyfall Research, on ServiceNow.
- 0.256System-change prediction scoreGPT-5.1 0.166 · Gemini 3 Pro 0.205 · Claude Sonnet 4.5 0.204
- $0Model spend570 predictions at zero calls; 84 on a frontier model cost $13.40
- ~80%Decisions answered from what it learnedYour AI took the rest
Ahead of GPT-5.1, Gemini 3 Pro and Claude Sonnet 4.5.
State prediction over three runs, compared case by case; the lead is significant against all three.
Step 1 of 2: The score
At predicting how the records change, Duplex scored 0.256, ahead of Gemini 3 Pro at 0.205, Claude Sonnet 4.5 at 0.204 and GPT-5.1 at 0.166.
Duplex first, your model behind it.
Duplex answers what it can and passes the rest to GPT-5.6.
Step 1 of 2: All to the model
Sending all 806 decisions to GPT-5.6 would cost about $129, every run.
How it was run.
- Benchmark
- World of Workflows, by Skyfall Research, built on ServiceNow: public, under the MIT licence.
- The work
- Predict how the records change after an operation, and work out which operation caused a change.
- Scoring
- The benchmark’s own scoring code, unmodified, checked against the authors’ published numbers.
- How Duplex ran
- Each step predicted before the answer was shown, learning only from steps already scored, with no model calls.
- Compared with
- The prediction files the benchmark’s authors published for GPT-5.1, Gemini 3 Pro and Claude Sonnet 4.5.
- Scale
- Three runs of 570 predictions, with paired significance tests on the same cases.
What it means for your work.
It learns how your systems behave.
Duplex predicted ServiceNow record changes better than three frontier models.
At no model cost.
570 predictions a run, with no model calls.
Your model where it helps.
Answering first with GPT-5.6 behind it, Duplex took about 80% of decisions and cut the bill from about $129 to $29.
See it on your own work.
From your own assistant to every agent your company runs, a pilot puts Duplex on your own work, with the measures agreed before it begins.