Two models.
One answer.
Every hard question goes into orbit around two models. One drafts, the other hunts for what is wrong with it, and the answer you read is the point they both settle on — with the whole trail one click away.
Draft, challenge,
resolve.
A single model grades its own homework. Dyad splits the job: the drafter commits to an answer, a model from a different family attacks it, and the resolver keeps what survived.
Draft
The drafter reads your prompt, files and context and commits to a complete first answer. No hedging, no self-editing: the goal is something concrete to test.
Challenge
A second model, chosen from a different provider, reviews the draft cold. It hunts for wrong facts, silent assumptions, missing cases and weaker alternatives.
Resolve
The challenger rewrites the final answer from both sides: it keeps the draft where it held, applies the fixes where it didn't, and says so when they still disagree.
Watch a pair
work a problem.
A recorded session. In the app every stage streams live, and you can open the draft and the review behind any answer.
Eleven models,
one set of names.
Pick any two, or let Auto Pair choose. Dyad names stay fixed while the engines behind them get upgraded, and every answer records the exact engine that produced it.
More ways to
use two models.
Pairing is the core. These are the tools around it: a blind test to find out which models you actually prefer, one-keystroke recipes, and answers you can hand to anyone.
Vote first, see names after.
Two models answer side by side on shuffled sides. You pick the better one, then the names are revealed and your personal ranking updates.
The right pair in one keystroke.
Press / in an empty box: code review, fact-check, decision memo, red-team. Each recipe sets the pair, the tools and the prompt for you.
Answers that show their work.
Every pair shows how many problems the challenger caught. Share the whole trail as one link — the content lives in the link, not on our servers.
Built for questions
that bite back.
For a quick fact, one model is enough and Dyad lets you use one. Pairing earns its extra seconds when a wrong answer is expensive.
Code & architecture
A second model reads the patch the way a reviewer would: edge cases, races, security.
try it → StrategyPlans & decisions
Get the case for, then the strongest case against, then a position that holds up.
try it → DocumentsReading & checking
Attach a PDF: one model extracts, the other checks the extraction against the source.
try it → MathProofs & numbers
Independent verification catches the arithmetic slip a single pass glides over.
try it →Chats stored on our servers. Your history lives in your browser; the server only relays the request you just sent.
Stages recorded per answer: which engine drafted, which challenged, what it cost, and whether you skipped the review.
Credits track real provider cost. Skip the challenge and you pay only for the stages that ran.
An economy for
verified work.
Credits pay for answers. $DYAD is the separate participation asset: staking, a share of the daily credit pool, and rewards for people who bring the network compute and reach.
Token design →Talk about it,
get credits.
Post about Dyad on X, paste the link, and the server verifies authorship and the mention on the spot. Verified posts are converted into credits instantly.
Start earning →Fair questions.
Isn't this just asking twice?
No. Asking the same model twice mostly gets you the same blind spots twice. Dyad pairs models from different providers, gives the second one a critic's brief instead of the original task, and only then writes the final answer.
How much slower is a pair?
Roughly two to two and a half times a single answer, since three stages run in sequence. Turn off the challenge stage to trade review for speed, or switch to single-model chat for quick questions.
Where is my chat history?
In your browser's local storage. Clear it and it is gone — we cannot restore it, because we never had it. Incognito threads are not saved at all.
Do I need the token to use Dyad?
No. Prompts are paid in credits. You start with 250, and you can earn more through Post & Earn. The token is for staking and network rewards, not for access.
Can I choose the models myself?
Yes. Auto Pair picks based on the prompt (code, images, long analysis, general), but you can pin any drafter and any challenger from the roster.