The Problem
Every AI gives you the same answer.
RLHF — the process that makes AI models helpful — also makes them predictable. Models learn to maximize the probability of the “best” response, collapsing a rich distribution of possible answers into a single safe, generic output.
Ask ten different models the same question and you'll get ten versions of the same answer. The creative outliers, the surprising angles, the weird-but-brilliant takes — they exist inside the model, but they're suppressed because their probability is too low.
OpenDrift brings them back.
How It Works
Verbalized Sampling in three steps
You send a prompt
Write any prompt. Choose a model. Set how many responses you want and how creative they should be.
The model samples its own distribution
Instead of returning one answer, the model generates multiple responses and self-assesses the probability of each one.
You explore the spectrum
Responses come back ranked from conventional to wild. Browse the full probability landscape and find what resonates.
See It In Action
One prompt, three realities
Pick a prompt. The top answer is what any chatbot gives you. The bottom one is what the same model was thinking but would never say first.
Prompt
gpt-4oWhat is the meaning of life?
The meaning of life is to find purpose through connection, growth, and contribution to something larger than yourself.
Life's meaning is an inside joke the universe tells itself — you're both the setup and the punchline, and the laugh track is entropy.
Meaning is a compression artifact. Consciousness evolved to lossy-encode reality, and 'purpose' is just the label we slap on patterns we can't unsee.
Illustrative examples. Your results vary by model, threshold and the mood of the universe.
Your Controls
Three dials. Infinite possibilities.
How many parallel responses to generate per prompt. Set it to 1 for a single alternative take, or crank it to 10 to explore the full spectrum. More responses means more chances to find something unexpected.
The creativity threshold. At 1.0, you get the most conventional responses. Drop it toward 0.01 and the model starts pulling from the deep tails of its distribution — the low-probability, high-surprise zone.
Each response includes a self-assessed probability score. This tells you how likely the model thinks this response is — letting you see exactly where it falls on the conventional-to-wild spectrum.
Use Cases
What will you drift?
Creative Writing
Get past the first-draft cliché. Surface metaphors, plot twists, and phrasings that a model would normally self-censor as too unusual.
Brainstorming
Generate ideas across the full probability spectrum. The best ideas often live in the long tail — the ones the model thinks are unlikely but you find brilliant.
Model Research
Compare how different models distribute creativity. Which model takes bigger swings? Which plays it safe? The probabilities tell the story.
Education
Explore how AI models think. See the full range of possible answers to a question, ranked by likelihood, and understand what RLHF suppresses.
Models
Compare creativity across providers
OpenDrift routes through OpenRouter, giving you access to frontier models from every major provider. Run the same prompt across different models and see who takes the bigger creative swings.
GPT-4o
OpenAI
Claude Sonnet 4
Anthropic
Gemini 2.0 Flash
Mistral Large
Mistral
Hundreds more on OpenRouter. Browse models and per-token costs.
The Spectrum
Every response has a personality
The response you'd expect. Safe, polished, and predictable. This is what every chatbot gives you by default.
Interesting territory. The model is reaching beyond the obvious — unexpected angles, fresher phrasing, genuine novelty.
Deep in the tail. These responses are unlikely, surprising, sometimes strange — and occasionally brilliant. This is where the magic lives.
What is Verbalized Sampling?
A prompting technique from Stanford and Northeastern researchers (2025). Instead of asking a model for one answer, you ask for several and have it state the probability it assigns to each. That one change recovers much of the diversity that alignment training (RLHF) suppresses, with no fine-tuning and on any model.
Isn't this just turning up the temperature?
No. Temperature reshapes how tokens are sampled and mostly adds noise. Verbalized Sampling changes what you ask for: the model reports where each answer sits in its own distribution, so you get a labelled range from safe to surprising instead of a lucky dice roll.
Are the percentages real probabilities?
They are the model's self-reported estimates, not token log-probs. They are consistent enough to rank answers and compare models, so read them as the model's confidence that it would have said this, not as ground truth.
Which models can I use?
Any chat model on OpenRouter: GPT-4o, Claude, Gemini, Mistral, Llama, DeepSeek and more. Run up to four side by side to see which provider takes the bigger creative swings.
What does it cost?
Free models run with a $0 balance. Paid models are pay-per-run at the OpenRouter price plus a 7% platform fee. There is no subscription and your balance never expires.
Can I use it from code?
Yes. The open-source @opendrift/sdk package and a REST endpoint give you the same ranked, probability-tagged responses in your own apps. Create an API key in the app and see the docs.