Free
€0forever, probably
For people who cannot decide yet.
Join the waitlist- 1,000 maybes per month
- Maybe 1 Mini
- Community support, sometimes
Maybe 1 is in private beta
Answers to every question, without ever being wrong.
Maybe is a language model trained on 41.7 trillion tokens of human hesitation. It reads your question, weighs every possible answer and replies with the only one that has never been proven false.
4,183 people on the waitlist. None of them has been told no.
Ask anything. Get a considered maybe.
Pick a question or type your own.
Teams currently thinking about it
- Perhapsly
- Kinda Capital
- eventually
- TENTATIVE
- Pending Group
- Somewhat
Confident answers are a liability
Most language models pick yes or no and hope for the best. When they guess wrong, somebody writes a postmortem and somebody else gets blamed.
Maybe never guesses. Since launch it has answered 2,906,114 questions, and not one answer turned out to be wrong. To be fair, not one turned out to be right either.
Typical model
Yes, definitely. Ship it on Friday.
Wrong 3.7% of the time
Maybe 1
Maybe.
Wrong 0.0% of the time
Yes is a risk. No is a risk. Maybe is a strategy.
Built for questions you would rather not answer yourself
One model, one word, no follow up questions from legal.
-
Zero hallucinations
Maybe has no facts, so it cannot make any up. Every reply is one word long and checked against itself.
-
Answers in 0.41 seconds
The answer is ready before you finish typing. It was ready before you started, to be honest.
-
Nothing anyone can quote in court
No advice, no promises, no liability. Our lawyers reviewed it and said maybe.
-
Unlimited context window
Paste 900 pages of contracts. Maybe will read none of them and answer with the same calm.
-
Fluent in 104 languages
Quizás, vielleicht, forse, может быть. The same answer with a local accent.
-
0
wrong answers since launch, out of 2,906,114 questions asked.
From question to maybe in three steps
Same steps in the playground, the API and the Slack app.
-
01
Ask anything
Strategy, medicine, relationships, deploys on a Friday. Any length, any language. Attachments welcome and left unopened.
-
02
Maybe weighs it
1.2 trillion parameters argue for yes and for no until the two cancel out exactly.
-
03
Get your maybe
One token at 50.0% confidence, delivered in 0.41 seconds. Same as last time. Same as next time.
Request
curl https://api.maybemodel.com/v1/answer \
-H "Authorization: Bearer $MAYBE_API_KEY" \
-d '{"question": "Should we migrate to Kubernetes?"}'Response
{
"answer": "maybe",
"confidence": 0.5,
"tokens": 1,
"latency_ms": 412
}Benchmarks we are proud of
10,000 questions from our internal evaluation set. We also wrote the evaluation set.
| Metric | Maybe 1 | Typical frontier model |
|---|---|---|
| Hallucination rate | 0.0% | 3.7% |
| Wrong answers | 0 | 412 |
| Right answers | 0 | 9,588 |
| Average answer length | 1 token | 611 tokens |
| Time to answer | 0.41 s | 8.7 s |
| Times it said “I apologize for the confusion” | 0 | 2,317 |
| Score on MMLU | Maybe | 88.4% |
Results may vary. They have not so far.
What early users are maybe saying
“We replaced our quarterly planning offsite with one API call. It took 0.41 seconds and nobody got blamed for anything.”
VP Strategy, Kinda Capital
“I asked whether we should raise a Series B. It said maybe. We raised a Series B. Technically it was right.”
Founder, Tentative Systems
“Support tickets fell 38.6% once customers learned every answer would be maybe. Churn went up a little. Maybe related.”
Head of Support, Perhapsly
Simple pricing. Same answer.
Every plan answers maybe. You are paying for how it feels.
Pro
Most hesitant€19per month
Same answer. Faster.
Join the waitlist- Unlimited maybes
- Maybe 1 and Maybe 1 Ultra
- Answers in 0.21 seconds
- Priority hesitation
Enterprise
Custom
Same answer, after a six week procurement process.
Join the waitlist- SSO and audit logs
- SOC 2 Type Maybe
- 99.95% maybe uptime
- An account manager who is not sure either
The never wrong guarantee. If Maybe ever gives you a wrong answer, we refund the whole year. Refunds paid so far: €0.00.
Questions, answered as clearly as we can
Which is to say, mostly.
Does Maybe ever say yes?
Maybe.
How accurate is Maybe?
Maybe has never said anything that was later proven false. We call that 100% accuracy. Our investors call it that too.
What was Maybe trained on?
41.7 trillion tokens of meeting notes, weather forecasts, terms of service and replies to wedding invitations.
Can I fine tune Maybe on my own data?
Yes. After fine tuning, Maybe still answers maybe, but in your brand voice.
Is Maybe safe?
Our red team spent six weeks trying to get a harmful answer out of it. They got 18,442 maybes and one resignation letter, which also said maybe.
Does it work in other languages?
In 104 of them. In Japanese it says たぶん, which is the same thing, only smaller.
Where do my questions go?
Nowhere. The demo on this page runs in your browser and never sends what you type to anyone. Details are in the privacy policy.
Will Maybe replace my job?
Only if your job is saying maybe. In that case, possibly. We are sorry, maybe.
Can I cancel anytime?
Yes. You can also not cancel. We support both.
When does the beta open?
Soon. We have been saying that since March, and we have not been wrong yet.
Ready to never be wrong again?
Join 4,183 people on the waitlist. We will maybe email you when your spot opens.
You are number 4,184 on the waitlist. We will maybe be in touch.