Plans the shortest route, with zero model tokens.
A small world of rooms, with the goal two to five moves away and a loop and a dead end on the way. Each task shows which move leads where, then asks for a route. Duplex, a reactive agent and GPT-5.5 got the same map.
- 100%Plan success, on every runA reactive agent: 0% · Duplex used no model tokens
- 0%A reactive greedy agent
- ~204Tokens per decision a frontier model spends to match it
Finding the shortest route.
Step 1 of 5: The world
The same answer, without the bill.
What GPT-5.5 and Duplex each paid for the same shortest routes.
Step 1 of 3: Tokens
Duplex spent no model tokens on any decision. GPT-5.5 spent about 204 on each.
The move that looks closest can lead nowhere.
Where each approach’s moves led, over ten runs.
Step 1 of 2: Reactive
A reactive agent’s first moves: about half went into the loop or the dead end, and none reached the goal.
Four checks, passed on every plan.
What was checked on every plan Duplex made.
Step 1 of 4: Shortest
Every plan was the exact shortest route.
How it was run.
- The work
- Find a route through a randomly generated world of rooms, two to five moves to the goal, with a loop that goes nowhere and a dead end.
- Scale
- One hosted run of 20 tasks, and ten independent runs of 12 plans each.
- Compared with
- A reactive agent that looks one move ahead, and GPT-5.5 given the same map in its prompt.
- Kept equal
- The same map and the same task for every approach.
What it means for your work.
Plans at no model cost.
Duplex matched GPT-5.5’s shortest routes with no model tokens.
Repeat decisions are quick.
Asked again, Duplex answered in a median 582 ms; GPT-5.5 took 1,647 ms.
It stops instead of guessing.
Started somewhere it had never been, it said so, every time.
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.