A measurement bench for curious minds
Slicebench helps you explore how AI outputs vary. Run a single prompt against up to 50 models, or run a dozen parallel follow-ups against one generated output. Slice how you want.
Sign in with a passkey. Bring your own OpenRouter key for paid runs. Get an OpenRouter key or create an account (opens in a new tab)
What makes a good decision?
Send the same prompt to each selected model.
Model A
Model B
Model C
Inspect repeated responses, with each model kept separate.
Open BenchMultifactorial Model Behavior Benchmarking
An assay is a structured experiment. Pick the setup that matches what you want to find out.
How much does the answer change?
Run the same prompt across models, compare different prompts, or repeat one request to see how stable the wording really is.
Measures patterns in text, not answer quality or correctness.
Open the BenchWhat can a small change set in motion?
Start with one question and one saved answer. Branch from that same answer with different follow-ups to explore sensitivity to wording, case and punctuation.
Branches are independent continuations, not a chain of answers.
Open CascadeWhat does a model inherit from the context?
Bring an existing conversation, choose where to stop its history, and compare how different models continue from the same point.
Apollo is the supported import format today. Assessments are your judgments, not automatic scores.
Open Inheritance AssayDoes the decision hold when the framing changes?
Supply your own evidence and questions. Repeat a native Jev request or vary its framing to inspect the returned probability distributions.
Jev tests your supplied decision setup; it does not judge other models’ answers.
Open Jev / DecisionsEvidence you can inspect
Keep the generated text, the returned model identity and execution receipts. Inspect individual responses and export your results.
Gauges carry an evidence tier and a plain-language explanation. Too little data gets a reason, not a made-up result.
Use your own OpenRouter key. Provider keys are encrypted before storage; decryption and provider requests happen in the private worker.
Slicebench is not production-ready. Evidence-tier thresholds are provisional, not statistically validated. Text gauges describe wording—not meaning, quality or correctness. Short samples can be insufficient, and English-prose gauges are unsupported for code or other languages, whatever tier a reading shows. Jev distributions and manual Inheritance labels are not proof of a correct decision.
Bench gauges include Zipf word-frequency slope, vocabulary variety, lexical density and most-used-word boards. Raw and grammar-word-filtered boards are shown separately. Models are kept separate rather than pooled into one average. Gauges are computed when a run executes and are not backfilled into older runs.
All model requests go through OpenRouter; this is not a comparison of direct provider integrations. Estimates are approximate, and reported costs can be partial or unavailable. Check your provider dashboard for billing truth. Review the estimate and paid-call consent in the launcher before starting an experiment.
Provider keys are encrypted at rest, not irrecoverable: the private worker decrypts them to make authorized requests. Recovery codes restore account access after new passkey enrollment. Save those codes when you create your account.
From curiosity to a recorded experiment
Choose a username, register a passkey and save your recovery codes.
Add an OpenRouter API key for paid experiments. Usage is charged to your provider account.
Choose a mode, review its inputs and cost estimate, then launch when you’re ready.