A random translation relay is funny exactly once. The languages are drawn for you, the result lands, you laugh, and then you want to know what would have happened if the Japanese stop had come before the French one. You can’t check. The dice are already gone.

Hypertranslate Route fixes that by handing you the itinerary. You pick the source language, you pick the final language, and you place every stop in between in the exact order you want. Then you pick one engine for the whole run and press Translate.

Same telephone game. Except this time you can rerun it.

What a translation route is

A translation route is the ordered list of languages your text passes through. The homepage and Hypertranslate AI both generate that list randomly and hide it from you until the run is over. A route is what you get when you write that list yourself.

That makes /route the fourth way to play, and the four modes differ in exactly two knobs:

The first two answer “what does randomness do to this sentence?” The fourth answers “what does a lot of randomness do?” Only the third lets you ask a question with one variable in it.

Why the order changes the result

Translation is lossy. Every hop throws away something the next engine can’t recover — a register, a connotation, the difference between “resign” and “step down.” Stack hops and the losses compound, because each engine is translating the previous engine’s guess, not your original sentence.

Order matters because the input to hop four is the output of hop three. Swap two stops and every downstream engine receives different text. The same six languages in a different order are not the same experiment.

Run English → Japanese → French → English and you get one sentence. Run English → French → Japanese → English and you get another. Neither is more correct. They’re different paths through the same lossy system, and the endings can be completely unrelated.

Which is the whole point of being able to set the order. A random relay shows you one path and destroys the evidence. A route lets you hold the path still and change one thing.

Step 1: Pick the two ends

Open Hypertranslate Route and start with the ends of the chain.

Source language is where the text begins. Leave it on Auto-detect and the engine identifies the input itself — useful when you’re pasting something you didn’t write. Set it explicitly when you want the first hop to be a real translation rather than a pass-through.

Target language is where the chain lands. This is the language you’ll read the final result in, so pick the one you actually understand. Everything between the two ends is yours to arrange.

Step 2: Add the stops in your order

The route preview under the form is the editor. There’s no separate settings screen — you build the chain where you’ll read it.

  • Insert anywhere. The dashed + slots between chips add a stop at that exact position, so you can drop a language into the middle of a chain without rebuilding it.
  • Change a stop in place. Every chip is a language selector. Swap one stop without touching its neighbors.
  • Remove a stop. The X on a chip deletes it and closes the gap.
  • Shuffle fills the route with random stops if you want a starting point to edit. Clear empties the middle and sends your text straight to the target.

The default route on load is Auto-detect → Japanese → French → English, so you can press Translate immediately and see the shape of the thing before you design your own.

You can stack up to 50 intermediate stops, which is 51 hops counting the final language. That ceiling exists for people who want to push a sentence until it collapses; most routes are funnier with five to eight stops, where the drift is still legible.

One detail worth knowing: if you’ve turned languages off on the language settings page, those languages won’t be offered as stops here either. Your language preferences carry across every mode on the site.

Step 3: Choose one engine for the whole route

Below the preview there’s a single engine selector. It offers Google Translate, Google Dictionary, MyMemory, and a list of AI models.

Pick one. That engine translates every hop of the route — no per-hop engine mixing, no silent switching halfway through.

That constraint is the feature. Google and a language model fail differently: Google maps phrases onto fixed equivalents and stays literal, while a model rewrites the whole sentence and holds idioms a little longer before losing them. If a run could switch engines mid-route, you’d never know which engine produced which damage. One engine per run means two runs are only ever one variable apart.

The engines also feel different to watch:

  • AI hops stream. The whole run travels over one WebSocket connection and each hop streams back token by token, so you watch the model rewrite your text in real time.
  • Google hops land in one piece. Fast, literal, and over before you can read it.

Your engine choice is remembered in the browser, so a comparison run doesn’t start with re-selecting it.

Step 4: Run it — and resume if a hop breaks

Press Translate and the route fills in hop by hop under “Your Route.” Each card shows its position in the chain (3/6), the language pair for that hop, and the translated text. The last card is flagged ★ Final · Result.

Every hop has its own Copy button, and so does the final result. If you only care about the moment the sentence died, copy that card and leave the rest.

Long routes collapse automatically to keep the page readable. You can expand all of them, collapse all of them, or open one at a time — the hop that’s currently translating opens itself, and the final result opens when the run finishes.

Things go wrong sometimes. A provider rate-limits you, a model hits its output limit, the socket drops mid-hop. When that happens:

  • Completed hops stay on screen. Nothing already translated is thrown away.
  • Press Continue and only the failed hop and the ones after it re-run.
  • Switch the engine first if you want. Continue resumes from the failed hop on the new engine, so a provider failure doesn’t cost you the whole route.
  • If you changed the route or the text, the tool tells you instead of quietly resuming into a different experiment — press Translate for a fresh run.

Finished runs are saved to local history with the exact language order and the engine used. Up to ten are kept, and replaying one restores every hop instantly without calling any API, because the texts are already stored.

Three routes worth building first

If you don’t have a route in mind, run these three. They’re designed to isolate one variable each.

1. Swap two stops. Build a six-stop route, run it, then move one middle stop two positions later and run it again with the same text. The two endings will differ, and you’ll know the order did it — not the languages, not the engine.

2. Same route, two engines. Keep the route fixed and rerun it once on Google Translate and once on an AI model. This is the comparison the single-engine rule exists for. Google tends to flatten idioms into literal equivalents; a model tends to keep the idiom’s intent one hop longer and then mangle it more creatively. Same itinerary, two different kinds of wreckage.

3. Grow the route one stop at a time. Start with English → Japanese → English. Add a stop, run again, add another. Watch the hop where the sentence stops being recoverable. That hop is usually earlier than people expect, and it’s different for every sentence.

Try it now

Head to Hypertranslate Route, build a route, and press Translate. Then change one stop and run it again — that second run is the thing no other mode on this site can give you.

If you want the theory behind why order matters so much, the 100-times meme covers where this game came from and why the drift is the point. If you’d rather let the site pick the languages, Hypertranslate AI does the dice-rolling for you on a language model.

And if your carefully planned route comes back as a sentence about livestock — you built that route. Own it.