Choir, online

started Nov 1, 2026 · last update Nov 2, 2026

The Choir desktop app, in the browser. Ask one question to every major AI model at once, watch the answers stream side by side and gather into groups by what they actually said, then carry the best ones into your next question. A second tab, the Collider, does the same for pictures: one prompt, several image models, side by side.

Sergio-Aragonés-style cartoon under a hand-painted banner reading CHOIR: seven colorful robots sing from sheet music on wooden risers, their songs streaming as glowing ribbons into two wooden corrals marked YES and NO; an owl conductor in a tuxedo at a podium peers through a magnifying glass and holds up a scorecard reading 8.4; on the right, three painters at easels paint the same red paper boat while an assistant juggles reference photos; a signpost reads choir-swirl.pages.dev.
OpenAI, Anthropic, Google, xAI, DeepSeek & moreA judge that scores and synthesizes Answers grouped by bottom lineBuild your own panelFiles & images in Image ColliderPay as you go, no subscription
Open the Swirl ▸ Open the Collider How to use it

The Swirl

Ask everyone. See where they agree, and exactly where they split.

  1. Pick a panel. The strongest model from each lab, a cheap-and-fast set, a red team, a reasoning panel, open-weights models, or your own. Up to 24 models answer at once.
  2. Watch it stream. Each model is a lane on a timeline: dotted while it thinks, glowing while it writes. Lanes whose answers share words drift toward each other as they go.
  3. Get judged. A judge model scores every answer, writes a synthesis, names the line they all agree on, and sorts the answers into corrals by their bottom line, with a note on where the minority veered.
  4. Carry forward. Pick the answers that ride into your next question: one card, the aligned ones, the dissent, a whole corral. The rest are cut from the map.
The Swirl after a question: five model lanes over time gather into two corrals, No, not sandwich (3 models) and Yes, sandwich (2 models), beside the answer cards and a synthesis.
“Is a hot dog a sandwich?” Five models, two corrals: three said no, two said yes, and each side says where it veered.
The reader open on the synthesis, with arrows to flip through every answer and a Compare button.
Flip through every answer with the arrow keys, or compare two side by side.
The Metrics tab: best answer, cheapest good answer, fastest first token, the cost of the turn, and each distinct answer with its members.
Metrics: best answer, cheapest good answer, fastest model, and what the turn cost.

Your own panel

The model wizard lets you browse every model with a key by price, context window and whether it can see images, then build a runset. Add the same model three times and give each copy its own instructions and its own frame for the question: an optimist, a skeptic who hears it from an investor, and a newcomer who hears it from a twelve-year-old. Same weights, three setups, one map.

The runset step of the model wizard: each voice has a name, running instructions with presets, a frame for the question, and an optional temperature.
One model, several voices: running instructions, a frame, and a temperature per copy.

Attach text, code, PDFs and images to any question; images go to the models that can see them.

The Collider

One image prompt, every image model, side by side.

  1. Break an image down. Drop in a reference and several vision models turn it into prompt pieces: subject, style, palette, composition, lighting, characters, what to avoid. Keep the best field from each.
  2. Keep bibles. Style, character and world sheets with their own reference plates, so a look or a face stays the same from render to render.
  3. Render everywhere. The same prompt to GPT Image, Grok Imagine and Gemini’s image models at once, each with its exact cost.
  4. Star, note, re-roll. Leave a note on a near-miss (“make it dusk, with fireflies”); the prompt is rewritten and the render goes along as a reference.
Three vision models' breakdowns of a red paper boat on a teal pond, each with Use buttons beside every field.
Claude, GPT and Grok each read the same picture; take the best field from each.
One prompt rendered by GPT Image 2, Grok Imagine and Gemini 3.1 Flash Image: a red paper boat and a sleeping frog on a pond, in three slightly different gouache styles.
One prompt, three image models, about 2 to 7 cents each.
A re-roll after the note make it dusk with fireflies: the same scene at dusk with tiny glowing dots over the water.
A note and a re-roll: dusk, fireflies, same boat.

What it costs

Choir online is pay-as-you-go: no subscription. One credit is a cent; models are charged at their public list price times two, per answer and per image, as they finish. Before a question starts it reserves its worst case and gives back what it didn’t use, so you never run past your balance.

PackCreditsRoughly
Taste · $5400a hundred questions to the top-model panel
Regular · $101,000about 250 questions
Value · $253,000 (+20%)about 750 questions

Rather pay the providers directly? The Choir desktop app does the same comparisons on your Mac with your own API keys, for free. Accounts are invite-only while the hosted version is in testing; ask me for one.

How it’s built

  • Choir’s own provider model. The same provider table, model registry and run format as the desktop app, so a turn exports as a Choir run that the app’s CLI and the Choir Reports scripts read as-is.
  • Cloudflare Pages and D1. Answers stream over server-sent events with an idle watchdog instead of a hard timeout, so slow thinkers get their minute and stalled ones get cut. The judge falls back to the next provider when one is busy.
  • Keys stay on the server. Provider keys are sealed in the database or set as deploy secrets and never reach the browser. Stripe handles payments; every charge and credit is one row in an append-only ledger.
  • Your stuff stays with you. History, profiles and Collider renders live in your browser and export as files. The server never keeps your images.

Built with Claude Code in two days, on top of the Choir macOS app and a lot of lessons from Codebackstory, the Scriptorium book pipeline and the Sticker Machine.