ChatGPT and Claude Together on iPhone: One Answers, the Other Checks
C1M (Duo) is a free button for your iPhone that makes ChatGPT and Claude work as a pair. You ask one question; ChatGPT answers, Claude reads that answer, checks it and fixes what is wrong, and about a minute later one checked answer is on your screen and in your clipboard. Duo runs on the ChatGPT and Claude apps already on your phone and on your own accounts (free ones are enough), through the "Ask ChatGPT" and "Ask Claude" actions those two apps publish for Apple's Shortcuts app. Nothing passes through a server of ours, because there isn't one. C1M is the open-source code behind Duo and its sibling Trio, which adds a third AI from a third company on a free key.
Why a second AI helps
An AI answers every question in the same confident voice, including the parts it gets wrong. In a 2025 study of more than 48,000 people in 47 countries by the University of Melbourne and KPMG, 66% said they rely on AI output without evaluating its accuracy, and 56% had made mistakes in their work because of AI (KPMG). The people who do check usually copy the question into a second AI and read two tabs side by side. Duo is that habit without the tabs and the copying. Google suggests "chatgpt or claude which is better"; Duo's answer is both, one writing and the other checking.
It pays off most on the question where being wrong is expensive: a contract clause, a supplier, a treatment you are about to start, a letter you can't unsend. For "what's the capital of France", one AI is plenty.
An AI from another company makes a good checker because it was trained differently. Microsoft builds the same pair into its own product: in the Critique feature of Microsoft 365 Copilot's Researcher agent, "GPT drafts, Claude reviews" (Petri, March 2026). The order matters too. In a July 2026 study of coding tasks, Claude reviewing drafts written by OpenAI's Codex raised the share of correct solutions from 71.6% to 89.7%, while Codex reviewing Claude's drafts lowered it from 91.4% to 82.8% (arXiv 2607.21656). Duo keeps the order fixed: ChatGPT answers, Claude checks. And when the two disagree, Duo shows you where, instead of blending them into one smooth paragraph. Two AIs agreeing is never presented as proof, since they read much of the same internet.
What one run looks like
The repository's demo shows a real answer copied from the phone. The question: "Should I switch to the cheaper supplier if the contract has a 30-day termination clause?" The person picks Critique, the default mode, which costs two messages.
The finished answer starts by saying how it read the question: "Reading chosen: the 30-day clause is in your current supplier's contract." Then the bottom line: a 30-day exit lowers the risk of switching, but it is a reason to weigh the move seriously, not to make it automatically. Then what to verify: notice formalities, early-termination fees, minimum commitments not yet met, discounts that could be clawed back. It ends with a confidence line ("medium — the direction is solid, but the decision depends on numbers and contract wording not provided") and a line on what Claude corrected in ChatGPT's answer; this time, nothing. When Claude does find errors, they are listed at the end, so you see exactly what was fixed. Then Duo offers "Done" or "Another mode — same question", which starts again with your question already filled in.
Twelve modes, one to start with
Critique is the only mode you need on the first day. The menu shows the price of every mode, in messages from your own ChatGPT and Claude plans, before you choose. Six everyday modes sit in the main menu (Critique, Debate until agreement, Advisor, Side by side, Synthesis and Auto); the rest wait one tap away under "More…". The other eleven include:
- Debate until agreement (3 to 8 messages): for life choices with no single right answer, such as "which one should I choose" or "should I agree". ChatGPT takes a side, Claude objects, ChatGPT answers the objections, for up to three rounds; the moment one side is honestly convinced, the rounds stop. Then Claude, as a neutral recorder, writes three short parts: where they agreed, what is still disputed and what is yours to decide, with the whole debate underneath.
- Advisor (1 message): Claude reviews your own text without rewriting it, strongest objection first.
- Side by side (2 messages): both answer separately, shown next to each other.
- Synthesis (3 messages): both answer without seeing each other, and one answer is the base while the other adds only what is missing. You pick the base: Claude for facts and structure, ChatGPT for tone and creativity.
- Auto (1 message plus the mode): not sure which mode fits? ChatGPT recommends one, and Duo starts again with your question so you can pick it.
- Clarify (2 messages): ChatGPT first asks what is missing, you answer in a small window, then Claude gives the precise answer.
- Image (2 to 3 messages): both draw the same description without seeing each other, and a page shows the two sketches side by side for you to pick. The drawing code lands in your clipboard, ready to turn into an image file of any size.
- Dispute map, Decision, Debate and Delta. A built-in help screen explains them all and costs nothing.
Trio: a third AI on a free key
Trio is Duo's Critique plus a third voice. After Claude's check, an AI from a third company reads the question and both answers and names only the errors and gaps both missed, in a separate block under Claude's answer; if there are none, it says so in one line. Claude's answer is already in your clipboard before the third voice is asked.
It runs on a free key: a long access code that NVIDIA, Google AI Studio, Groq or OpenRouter give out at no cost and without a card. NVIDIA's free models are asked first (Kimi K3, GLM-5.3, DeepSeek V4.1 Flash), then Gemini Flash on Google AI Studio, Qwen and gpt-oss on Groq, and OpenRouter's free models. NVIDIA's sign-up asks for an email, a password and an SMS code; Google AI Studio and Groq need no phone. On OpenRouter, switch on free models once in its privacy settings. You can save several keys, one per line, and Trio walks every model of every key until one answers: a busy model passes the turn to the next one. The free limits are counted per account, so when one company says "limit reached", Trio moves on to the next company for that run; extra keys help most when each comes from a different company. If you already pay for an AI service, Trio also takes your own paid key from xAI (Grok) or DeepSeek, asked before the free ones, or from Anthropic or OpenAI, asked after them; you pay that company directly.
The keys live in a file in your own iCloud Drive (Shortcuts/poly-key.txt), not inside the shortcut, so sharing Trio with a friend never shares your key. To add a key later, paste it on a new line in that file. If every key and model fails, Trio tells you in plain words what each one said (no money left, limit, overloaded, wrong key) and what to do next.
How to start
You need an iPhone on iOS 18 or later with the ChatGPT and Claude apps installed and signed in. Free accounts are enough: Duo was run end to end on free ChatGPT and Claude accounts on 1 October 2026, and free Claude runs Sonnet 5.5, the same Sonnet as the paid plan. Each app answers with the model set as its default, so a glance at the model in the Claude app before an important question is a good habit. Have only one of the two apps? Start by installing the other; have neither? Gemini or Meta AI can walk you through installing both, free.
- On the iPhone, open the table in the README or polyhelper.ai/duo, and tap Duo in your language: English, Spanish, Portuguese, Russian or Ukrainian. The Shortcuts app opens; tap Add.
- Run it once and tap Always Allow each time it asks about ChatGPT, Claude and the clipboard (four to six times, on the first run only).
- Type a question, pick Critique and leave the screen alone for about a minute, so the run can finish. Multi-step modes take closer to two.
The signed Duo.shortcut and Trio.shortcut files are also in the repository's releases/ folder and in the v1.0.0 GitHub release; keep the name "Duo", since "Another mode — same question" finds the shortcut by it. The answer comes back in the language of your question. Shorter questions come back most reliably; for a long text, the Compress companion below shrinks it first.
Free Claude has a daily limit. When you reach it and want more, Claude Pro ($20 a month, $17 billed yearly, paid to Anthropic) raises it at least five times; in the meantime Trio's free third voice keeps checking.
Everywhere on the phone
Duo starts from anywhere a shortcut can: select text in any app and share it into Duo, with your question box already filled; press the Action Button; double-tap the back of the phone (Back Tap); tap a Home Screen or Lock Screen widget or a Control Center button; touch an NFC tag on your desk; or say "Hey Siri, Duo". The final answer also goes to the shared clipboard, so on a Mac or iPad nearby you paste it with Cmd+V.
Each run is saved to a plain-text journal in your own iCloud Drive, with the drafts of both AIs, so nothing is lost if an answer gets cut off; the free 5 GB of iCloud covers it, and without iCloud the answer still reaches your screen and clipboard. At most once a week Duo checks for a newer version and offers it in one line.
On Android, and with the Poly A1 app
The Poly A1 app (free to download; Duo and Trio stay free in it) installs Duo and Trio for you and walks you through every step. On Android it runs the same pair automatically: one strong free AI answers and one from another company checks, on your free key, NVIDIA's models first, then Google AI Studio, Groq and OpenRouter. That Android Duo was run on an Android emulator on 1 October 2026.
Open and checkable
The prompts are plain text files in locales/en/prompts/, every word on screen comes from locales/<lang>/, and tools/verify.py checks that the signed shortcuts are exactly what the sources produce. The repository's evidence page lists 27 sources from 2025 and 2026, including studies where pairing did not help, and maps each rule in the prompts to its source. The same build also makes small companions: Voice (dictate, hear the answer read aloud), Photo (share a photo or PDF, the text is read on the phone), Quiet (a Critique run with no screens, straight to the clipboard), Compress (Apple's on-device model shrinks a long text first, on Apple Intelligence iPhones) and Multi (the pair plus any other AI apps you have, merged into one document).
C1M is written by Vadym Chernets, released under the Apache-2.0 licence and archived on Zenodo (DOI 10.5281/zenodo.23089503) and Software Heritage. On a computer, the same idea runs in three Claude Code plugins: Duocall asks one AI from another company before a decision, Sidecall asks the AI programs installed on your computer, and Roundcall asks AI chat websites in your browser. The code is at github.com/vadimchernets/c1m-duo.
Vadym Chernets also builds Poly A1, an app in which AIs from different companies check each other's work.