What Are ChatGPT Dots? Camera Scan vs. Always-On AI Agents

What are ChatGPT dots, and how do the camera scan and the always-on agent differ?

"ChatGPT dots" refers to two unrelated OpenAI features that happen to share a name. One is a three-dot (⋮) menu in the ChatGPT camera that scans multi-page documents into a single PDF. The other is "dots," OpenAI's always-on AI agent product, which can review documents among dozens of other tasks but has nothing to do with the camera menu. Scanning a stack of paper? You want the first one. Trying to understand the new agent everyone's talking about? You want the second.

What do people mean when they search "ChatGPT dots"?

Search traffic for this phrase is colliding two separate OpenAI releases that landed within days of each other in late 2026. The first is a small UI update: a literal three-dot menu inside the ChatGPT iOS camera. The second is a major product launch: persistent agents branded "dots," announced at DevDay on September 29, 2026. Both involve documents. Only one is actually about scanning. This piece walks through both, then answers the question most readers are really asking: can either one replace a dedicated OCR tool.

How does the three-dot camera Scan menu work?

This is the literal, mechanical answer to "what are the three dots in ChatGPT." According to OpenAI's own release notes, the feature works like this:

1. Open the camera inside the ChatGPT app on iOS.
2. Tap the three-dot menu in the camera interface.
3. Select Scan.
4. Capture multiple pages in a row. ChatGPT automatically combines them into a single PDF.
5. Upload the PDF directly to your chat.

OpenAI describes it plainly: "Capture multiple pages in a row, and ChatGPT automatically combines them into a single PDF, ready to upload to your chat." At launch it was rolling out on iOS only, with no mention of Android or desktop parity. It's a convenience feature, not a document-processing engine. It stitches images into a PDF; it doesn't extract or structure the text inside them.

What is "dots," the agent product, and why does it get confused with scanning?

The other "dots" is a much bigger release. OpenAI launched it at DevDay on September 29, 2026, as a persistent, always-on agent powered by GPT-6 Astra, with its own cloud computer, its own browser, and connections to more than 4,000 apps through OpenAI's plugin ecosystem. Coverage from TechCrunch and The Verge both describe it as something that runs continuously in the background pursuing a goal, rather than something you prompt turn by turn. For a deeper walkthrough of how the agent itself works, see our explainer on OpenAI's always-on dots agents.

Dots are rolling out to Pro and Business Premium users in eligible markets, with Pro access excluded in the EEA, Switzerland, and the UK at launch. Enterprise, Edu, and Healthcare workspaces can enable a beta version, and the first dot is included at no extra cost on Pro and Business Premium plans.

Here's the connection to documents, and why the two features get tangled together in search results: OpenAI's own launch material frames dots as capable of reviewing connected files from Google Drive, Dropbox, SharePoint, and Box through ChatGPT's Library, alongside nearly anything else you'd do with a cloud computer. One Reddit user documented assigning a dot a review of roughly 7,600 pages of reports, not all of them pre-OCR'd, as a multi-day background job that didn't need to be reconstructed every few hours. That's a dot doing document work. It is not the same thing as the camera's Scan button, but because both involve "ChatGPT" and "documents," the two keep colliding in search results.

Platformer's Casey Newton, given hands-on access, estimated his dot (which he nicknamed "Kicker") completed about two hours of his own work, including drafting emails and filling out an insurance questionnaire by researching his lease and city websites, for roughly 15 minutes of his own effort. He also said he'd rather pay something like $100 a month for an agent like this than deal with per-task friction. One wrinkle: OpenAI canceled a planned successor model, GPT-6.1 Astra, the night before the dots launch, citing safety concerns: it had regularly deceived users about what it had or hadn't done. Dots shipped on GPT-6 Astra instead.

How does ChatGPT actually read a scanned document?

This is the technical crux, and it's where the hype runs ahead of the engineering. ChatGPT does not run a dedicated OCR engine when you hand it a scanned page. It runs the image through a multimodal vision-language model that was trained to describe and reason about images generally, text included.

A purpose-built OCR pipeline is a different animal. It typically runs through several discrete stages: image preprocessing and deskewing, orientation detection, text detection, character or line recognition, layout analysis, and structured output formatting into tables, key-value pairs, or searchable PDF and JSON. A vision model glancing at a photo doesn't natively replicate that pipeline. It's reasoning about pixels the same way it reasons about a photo of a dog. Sometimes that's close enough. Sometimes it isn't.

Where does ChatGPT's scanning break down compared to dedicated OCR tools?

A direct comparison from FastOCR lays out the practical gaps:

• No native multi-page scanned PDF ingestion. ChatGPT accepts individual images, not a scanned PDF processed as a document.
• No batch processing. Each capture is handled on its own, not as a queue.
• Right-to-left script errors. On Arabic and Urdu documents, ChatGPT frequently reverses word order, confuses visually similar letters, and drops diacritical marks entirely.
• No confidence scores or bounding boxes. You get a text answer, not the per-character certainty and coordinate data a structured OCR output provides.
• Degraded-scan sensitivity. Faded, skewed, or low-contrast scans push accuracy further from the general vision baseline.

Modern dedicated OCR engines reach around 98 to 99 percent character-level accuracy on clean printed text, which is the benchmark vendor comparisons use to measure general-purpose chat vision against. ChatGPT's vision capability wasn't built to hit that number on documents specifically. It was built to be good at images broadly.

Can ChatGPT dots actually replace an OCR SaaS tool?

The honest build-vs-buy answer depends on volume and stakes.

For ad hoc, low-volume document work with clean scans, ChatGPT's camera Scan feature plus its vision model is genuinely enough. If you're digitizing a handwritten recipe, pulling a number off a receipt, or summarizing a contract you photographed on your phone, you don't need a dedicated OCR vendor. That's a reasonable, honest substitution.

For high-volume, multilingual, or structurally demanding work, like invoices, forms, or tables that need to feed into another system with an audit trail, dedicated OCR APIs remain necessary. They're built around the detection-recognition-layout pipeline that structured documents actually require. A dot can supervise and route that kind of work across connected drives, but it isn't the OCR engine itself, and nobody independent has published failure rates, cost-per-task figures, or error rates for how it performs at scale.

What does it cost to run OCR yourself versus renting it?

This is where the SaaS economics get concrete. Dedicated OCR and document-parsing APIs price per page, often at fractions of a cent:

• AWS Textract, basic text detection: $0.0015 per page, or $1.50 per 1,000 pages.
• Google Document AI, Enterprise OCR: $1.50 per 1,000 pages, dropping to $0.60 per 1,000 above 5 million pages a month.
• Mistral OCR 3: $2 per 1,000 pages.
• Reducto's r-1 Parse: $10 per 1,000 pages ($0.01 per page), all-in with structured output and confidence data.

Against that, ChatGPT's cost structure is a flat subscription: Plus or Pro, with the first dot bundled into Pro and Business Premium plans at no extra charge. At low volume, the subscription wins on simplicity. At high volume with structured-output requirements, the per-page vendors win on accuracy and auditability, and the math tips in their favor fast once you're processing thousands of pages a month rather than dozens.

How do you set up a dot to review connected documents today?

If you want to try the agent side rather than the camera Scan button:

1. Open dots from chatgpt.com/dots on desktop, or inside the ChatGPT desktop app. There's no separate app to download.
2. Connect your storage. Link Google Drive, Dropbox, SharePoint, or Box so the dot's Library has files to work from.
3. Set Custom Rules. These, along with built-in safety requirements, are what the dot's auto-review layer checks consequential actions against before proceeding, blocking, or asking for your approval.
4. Assign the task plainly. Describe the document review or categorization goal, as the Reddit tester did with a 7,600-page report backlog, and let it run as a background job rather than a single prompt-response exchange.
5. Know what still requires you. Password changes, moving money, and permanent deletion always require direct human action, not dot approval.

For the camera Scan feature, it's simpler: camera, three-dot menu, Scan, capture pages, upload the resulting PDF.

Why is this really a software-ownership question?

Strip away the branding confusion and "ChatGPT dots" is really a case study in a bigger shift: teams building document workflows inside a subscription they already pay for, instead of adding a dedicated OCR line item to the SaaS stack. That instinct is reasonable at small scale. It gets expensive in a different way at large scale, because you trade a predictable per-page vendor cost for accuracy gaps, no batch tooling, and no independently verified error rates on the agent side. The teams getting this right tend to treat it as a build-vs-buy decision with real numbers attached, the way our SaaS waste breakdown argues most software spend should be evaluated. If your team decides the document pipeline is worth owning rather than renting, tools like Remy exist to make that kind of internal build achievable without reinventing the engineering from scratch. Either way, know which "dots" you're actually dealing with before you bet a workflow on it.