TL;DR
ChatGPT Vision (GPT-5.5, the default model since April 23, 2026) reads screenshots, charts, and handwriting well — until it doesn’t, and then it misreads a number, drops a row, or invents a label that was never on the chart. The fix is not to stop using Vision. It is to (1) crop tight and upload at full resolution, (2) tell the model exactly which read you want (transcribe vs. explain vs. extract), (3) use a prompt that forces it to mark anything unreadable as [?] and to stop rather than guess, and (4) spot-check the values that actually matter. On dense charts and tiny text, switch the model picker to GPT-5.5 Thinking before you upload.
When Vision earns its keep
- Transcribing a screenshot where retyping would take 10+ minutes — a tax form, a long error log, a printed table.
- Explaining a chart, diagram, or UI flow to a colleague who needs a written summary.
- Triaging an error dialog to find the one actionable line buried in stack-trace noise.
- Turning a whiteboard photo of handwritten notes into a structured outline.
If you only need “describe this image artistically,” none of this applies. This guide is for facts that have to be right.
What ChatGPT Vision can take (as of June 2026)
Vision is built into the GPT-5.5 family — there is no separate “Vision model” to switch on. Upload an image in any chat and the model reads it. The platform caps, not the model, decide how much you can push through:
| Limit | Free | Plus ($20/mo) | Notes |
|---|---|---|---|
| Images per day | ~2 | ~50 | Free is tight; Plus also has an 80-files-per-3-hours rolling window |
| Max size per image | 20 MB | 20 MB | Same cap on both tiers |
| Per-chat upload total | 512 MB | 512 MB | Across all files in one conversation |
| Formats | PNG, JPEG, WebP, static GIF | same | Animated GIF and most raw camera formats are rejected |
Caps are platform-level, so moving from the default model to GPT-5.5 Thinking does not raise them. Numbers above reflect OpenAI’s published limits as of June 2026 and change without notice — confirm in the OpenAI File Uploads FAQ.
Before you upload
- Crop to the part you care about. Vision tiles a high-res image into 512×512 segments, so a full-screen capture spends attention on the menu bar and tabs instead of the one window you need.
- Zoom small text before screenshotting. OCR fails most often on text under roughly 12px — logs, code, footer fine-print. Re-export from the source app at full resolution rather than grabbing a chat-app preview.
- Pick the read type first. “Transcribe” (verbatim), “explain” (interpret), and “extract” (structured fields) have different prompts and different failure modes. Asking for transcription and analysis in one prompt degrades both — do it in two passes.
- Switch to GPT-5.5 Thinking for dense images. For multi-region charts, packed tables, or low-contrast scans, the Thinking variant does the cross-region comparison the default model rushes. Pick it in the model picker before you send.
Step by step
-
Upload the image and wait for the attachment to actually finish. Sending the prompt before the upload completes silently drops the image.
-
State the read type explicitly. For verbatim text:
Transcribe the visible text in this screenshot verbatim. Preserve line breaks. Mark any character you can't read clearly as [?]. Do not infer text that isn't visible. -
For charts, ask only for what is verifiable (axes, labels, trend shape) and forbid estimated values:
Describe this chart: axis labels, title, legend, and the shape of each series. Do NOT estimate specific y-values from the visual — only report values that are explicitly labeled on the chart. -
For UI screenshots, ask the model to walk the visible actions in order:
This is a screenshot of a settings page. List every action the user could take here, in the order they appear on screen. Don't infer behavior — only describe what's visible. -
For handwriting, expect ambiguity and ask it to expose its guesses:
Transcribe these handwritten notes. For any word you're uncertain about, give your best guess followed by [?] and 1-2 alternative readings. -
If the model says it can’t read a region, re-upload a higher-resolution crop. Don’t let it guess.
The prompt that surfaces uncertainty instead of hiding it
Vision task: [transcribe | explain | extract] this image.
Constraints:
- Anything you can't read with confidence, mark with [?].
- Do not invent text, numbers, or labels not visible.
- If a chart value isn't explicitly labeled, do not
estimate it from the visual.
- If the image is too low-resolution for the task,
say so and stop instead of guessing.The “say so and stop” clause is the one that matters most. Without it, the model produces a plausible-looking output instead of telling you the image was unreadable.
For high-stakes extraction, OpenAI’s own document-understanding guidance points the same direction: ask for strict JSON with null for any field you can’t read, and add an Evidence column that quotes the exact pixels the value came from. Models rarely hallucinate a field when the schema forces them to either ground it or leave it null:
Extract these fields as JSON: invoice_number, date, total.
Use null for any field you cannot read with confidence.
Add an "evidence" string for each field quoting the exact
text you read it from. Do not fill a field from context.Quality check
- Verbatim transcription: spot-check 3-5 lines against the image. If any are wrong, re-prompt with “be more careful with line N” or upload a higher-res crop.
- Charts: confirm the trend description matches what you see. For any specific number returned, verify it against an explicit label on the chart.
- UI explanations: confirm every button or field mentioned is actually visible. Phantom buttons are a known failure mode.
- Handwriting: read the original yourself for any critical line. The model handles printing well, cursive less well, and abbreviations badly.
Common mistakes
- Uploading a full-screen screenshot when you only care about one window — Vision gets distracted by chrome around it.
- Asking “what does this chart say?” with no constraints. It reads the trend and invents specific numbers that aren’t labeled.
- Trusting Vision on small text. Logs, code, and fine print are where OCR errors silently slip in.
- Letting the model fill in fields that are cut off at the screenshot edge — it produces plausible completions that don’t match the original.
- Using a low-resolution preview from a chat app. Re-export from the source at full resolution.
- Combining transcription and analysis in one prompt. Split them.
FAQ
- Which model reads images — do I need to turn Vision on?: No. Vision is built into the GPT-5.5 family, so any chat can read an upload. For dense charts or tiny text, switch the picker to GPT-5.5 Thinking before uploading; it does more careful cross-region comparison than the default.
- Can ChatGPT read handwriting?: Printing yes, neat cursive usually, messy cursive poorly. Surface uncertainty with
[?]and verify critical lines yourself. - Why did it invent a column that’s not in my chart?: Usually the chart was small or the legend was cropped. Re-upload a cleaner crop and constrain the prompt to “only what is explicitly labeled.”
- Does Vision work on math equations?: Reasonably for printed equations, poorly for handwritten ones. Ask it to render the equation in LaTeX so you can compare it back to the source.
- How many images can I upload per day?: As of June 2026, roughly 2/day on Free and ~50/day on Plus, with a 20 MB per-image cap and a rolling 80-files-per-3-hours window on Plus. Limits change — check the OpenAI File Uploads FAQ.
- Is it OK to upload screenshots with personal data?: Check your plan’s data-retention settings and redact identifiers first. As a rule, anything you wouldn’t paste into a normal chat shouldn’t go in as an image either.