Decision models arena

PAC-MAN ARENA

Frontier models think slowly and expensively. Local Jev-style decision models answer in milliseconds for fractions of a cent. Who survives longer?

Free, no signup · runs continuously · want hundreds of experiments on real data? Join the beta.

FRONTIER

OpenAI · ~1400ms/decision · $0.00420/decision

TIME
—
COST
—
SCORE
—
STATE
idle
LOCAL JEV

Ollama · ~35ms/decision · $0.000004/decision

TIME
—
COST
—
SCORE
—
STATE
idle

Beyond the simulation / beta

TEST YOUR CONFIGURATION ON REAL TRAFFIC

Pac-Man makes the trade-off visible. Your actual traffic tells you what works. Pick one change, compare it with your current configuration on the same kind of real sessions, and check quality before claiming any time or cost savings.

ONE CHANGE AT A TIMESame traffic type. Repeated runs. Verified result.
How an experiment worksInspect a session, classify its work, decide on one change, verify baseline and changed runs three times each, then compare and keep or drop the change.STEP 1Inspectread the sessioncost · size · toolsSTEP 2Classifywhat kind of work?long · coding · searchSTEP 3Decidepick one changee.g. cheaper modelSTEP 4Verifyre-run it both waysdid it finish the job?STEP 5Compareagainst the baselinecheaper? as good?baseline × 3with change × 3POSSIBLE RESULTkeep itdrop itSteps 1–3 only read data and cost nothing.Steps 4–5 run in a sandbox and cost money.
Every experiment follows the same five steps. Read a real session, sort it by the kind of work, pick one change that could make it cheaper, then re-run the task in a sandbox: once unchanged (the baseline) and once with the change, three times each. Keep the change only if it is cheaper and does the job as well. The re-runs are compared with each other, not with the original session, because a re-run is usually shorter than the real thing. Real-traffic experiments require beta access and a connected workload; none are running on this page yet.

Explore / 08 experiments

EXPERIMENT CATALOG

Select an experiment to see what changes and how it would be checked.

Ready to test your own configuration against real traffic?

Simulation: each model's decision quality, latency and per-decision cost are modeled profiles, not live API calls.