AI Tinkerers Prague 2026

Join
cheaperuh
party.

Čipera means lively, brisk, and nimble in Czech. Here it is affectionate sarcasm for AI that is too eager to think and spend. A reasoning router should choose the smallest useful amount of thought.

View the project on GitHub

Illustration of reasoning levels and agent orchestration
222 / 243supervisor completions
91.4%supervisor success rate
$0.0246combined supervisor cost

Measured OpenRouter calibration: 33 candidate tasks, 27 selected tasks, three supervisor choices, and three worker attempts per choice.

The supervisor chooses freely

The supervisor sees a vague brief, never the hidden answer. It may choose any OpenRouter-supported effort, including zero reasoning, then the worker makes three attempts at that choice. Three independent supervisor choices create the other matrix axis.

nonezero reasoning
minimalsmall check
lowshort chain
mediumbounded work
highdependencies
xhighdiagnosis
maxlast resort

All seven levels are selectable. The benchmark rewards choosing the lowest safe one, not choosing a fashionable level.

The benchmark in one minute

Vague briefSupervisor cannot see the answer.
One choiceAny level, including none.
3 x 3 evidenceThree choices, three worker attempts each.

A good result is not simply high accuracy. It matches an always-high control while using less combined worker and supervisor cost.

Measured three-arm comparison

Each strategy made 243 canonical worker attempts. Cost includes the supervisor's own selection calls where applicable.

StrategyVerifiedWorker tokensWorker costSupervisor costCombined cost

3D task frontier

Orange X: task difficulty0 is easy. 1 is hard.
Green Y: supervisor effortnone to max selected effort.
Blue Z: worker tokensmeasured mean token cost.
low

Loading measured task data.

Drag to rotate. Click a point for its measured result.

    Orange X is audited difficulty. Thick orange rings separate overlapping green selected-effort task dots. Blue dots are the other measured tasks.

    All 27 task outcomes

    Every row is a three-choice by three-attempt matrix. Supervisor selections and rationales are published without hidden prompts or model responses.

    Category and taskSupervisor choicesSupervisorAlways noneAlways highSupervisor cost

    Download the complete sanitized evidence: expanded-v1.json.

    Run the full calibration

    Run the free-choice supervisor beside always-none and always-high. The report logs worker tokens, supervisor overhead, and combined dollars per task. One paid unit is one decision plus three worker attempts.

    npm run benchmark:live -- --strategy=supervisor --run=my-run --limit=81

    npm run benchmark:publish -- --run=my-run

    Keys remain in ignored .env; no key is shown in output or checked into Git. Model details: OpenRouter model page and OpenRouter reasoning guide.