Paste a chart into ChatGPT or Claude and you get a read in seconds. Whether that read is right depends on something the vendors’ own docs are specific about: how an image gets resized before the model ever looks at it. Here is what vision models get right and wrong on a chart, our own blind test on a real candlestick chart, what a structured read covers instead, and how Ethy’s four-timeframe chart read works.
November 16, 2026
13
min to read
The short answer
Vision models describe a chart’s shape well, but they estimate prices from pixels, not from the data behind them, and the vendors’ own docs admit they struggle with exact positions and counting. In our blind test on a real candlestick chart, Claude got the high, the low and their dates about right, but overestimated how many candles were red.
It covers two different things that get asked for in the same sentence. The first is pasting a screenshot into ChatGPT or Claude and asking what it shows: both can take an image directly in a chat, and both will describe a trend, call out a level and name a pattern. The second is a structured read: handing the model, or a tool built for the job, the actual price and indicator numbers, so a question like “where is support” is a lookup in real data instead of a guess from a picture.
The two paths fail differently. A screenshot read is only as good as what the model can see in a picture that may get resized before it ever looks, which the next two sections cover in detail, with a real test. A structured read is only as good as the data and the rules behind it, which is where our library of chart and research prompts and the structured-read section below come in. Claude connected directly to TradingView’s own charts and data is its own setup, covered in how to connect Claude to TradingView with MCP.
Both vendors publish the same kind of warning, in their own words. OpenAI’s vision guide says its models “may struggle to understand graphs or text where colors or styles… vary,” which describes a candlestick chart almost exactly: solid bodies, thin wicks, two colors carrying the only signal of whether a bar closed up or down. The same guide adds that the model “struggles with tasks requiring precise spatial localization,” and separately that images “may be resized before analysis.”
Claude’s vision docs are just as direct in their own Limitations section: “Claude might hallucinate or make mistakes when interpreting low-quality, rotated, or very small images under 200 pixels.” On reading an exact point: “Claude’s coordinate and localization outputs are approximate… verify outputs before relying on them.” On tallying anything across a busy image: “Claude can give approximate counts of objects in an image but might not always be precisely accurate, especially with large numbers of small objects.” The section ends bluntly: “Always carefully review and verify Claude’s image interpretations, especially for high-stakes use cases.”
The mechanism behind this is specific, not vague. Claude’s coordinates guide explains that it reads images in 28-by-28-pixel patches, and a model on the standard resolution tier caps an image at 1,568 pixels on its long edge and 1,568 of those patches in total. A plain 1920x1080 chart screenshot, the guide says, “resizes to 1456×819, not 1568×882,” because the patch budget, not the edge length, is what forces the resize. Claude 4.7 and later models sit on a high-resolution tier, 2,576 pixels and 4,784 patches, which keeps that screenshot at full size, but the vision docs show a 4K one still shrinking to 2576x1449. Thin axis gridlines and small price labels are exactly the detail a resize like that loses first, before the model has formed any opinion about the chart at all.
Independent research backs up the general pattern on charts. CharXiv, a mid-2024 benchmark built at Princeton and other universities on real charts pulled from the wild rather than templated ones, found that its best-scoring proprietary model at the time answered only 47.1% of its reasoning questions correctly, against 80.5% for humans, and, in a stress test on older benchmarks, that “slightly different charts or questions can deteriorate performance by up to 34.5%.” The paper’s authors built CharXiv partly because older benchmarks, in their words, “often focus on oversimplified and homogeneous charts,” leading to “an over-optimistic measure of progress”; they fault the widely used ChartQA for the “homogeneous appearance” of charts drawn from only four websites. Neither benchmark is built around candlestick trading charts, and both are now over a year old, so treat the exact score as a snapshot of 2024 models on general charts, not a verdict on what today’s models do on a price chart. One benchmark that is built around them, MME-Finance (November 2024), tested 19 multimodal models on financial images, from screenshots to phone photos of charts, and reported that their performance was “particularly poor in categories most relevant to finance, such as candlestick charts and technical indicator charts.” The best closed model it tested, GPT-4o, scored 63.18 out of 100.
Rather than stop at what the docs warn about, we tried it on one real chart and checked the answers against the data. Hyperliquid’s public API gave us 60 daily candles of its BTC perpetual, September 18, 2026 to November 16, 2026, taken at 09:16 UTC on November 16, 2026, so the last day was still open. We drew them as a plain candlestick chart, dark background, green and red bodies, a price axis labeled every $2,000, date ticks every ten candles. Then we showed that image, and nothing else, to Claude Sonnet 5 in a fresh session with no access to the data, told it not to analyze the pixels with code, and asked it five questions on November 16, 2026. Here is what it said, next to what the data says.
| What we asked | What the image-based read said | What the data actually says |
|---|---|---|
| The highest point on the chart | About $87,000 to $87,500, in the first fifth of the chart, between Sep 20 and Oct 1, with two early spikes of almost the same height (medium confidence) | $87,471, on September 21, 2026; the next-highest high, $87,333 on Sep 23, was $138 lower |
| The lowest point on the chart | About $62,000 to $63,000, a long lower wick a few candles from the right edge, after the Nov 7 tick (medium-high confidence on the price, medium on the date) | $61,970, on November 15, 2026, the second-to-last candle |
| Color of the last candle (Nov 16) | Red, with a small body (medium confidence) | Red, with a $1,086 body, as of 09:16 UTC with the day still open |
| Color of candle 30 of 60 (Oct 17) | A small red body in a choppy stretch between $82,000 and $84,000 (low confidence: it had to estimate which candle was the 30th, and said it could be off by one or two) | Red, with a body of only $138 |
| Green candles vs red, across all 60 | More red than green: roughly 33 to 35 red and 25 to 27 green, from scanning the chart in segments rather than counting one by one (medium confidence) | 29 green, 31 red, a gap of 2 |
The extremes held up. Both price answers were ranges $500 to $1,000 wide: the high fell inside its range, the low just $30 below it, and both landed in the right stretch of dates. The last candle’s color was right too. So was candle 30’s, but at the low confidence the reader gave it, and two of the four candles within two places of it are green, so the miscount it warned about could easily have flipped the answer. The count is where it slipped. It got the direction right, more red than green, but put the gap at 6 to 10 candles when the real split was 31 to 29: its red estimate was 2 to 4 too high, and its green estimate 2 to 4 too low. One plausible reason is size, not number: in the last 30 candles, the 16 red bodies add up to about $27,800 from open to close, against about $10,600 for the 14 green ones, so red dominates the picture while the count stays close. That fits the failure Claude’s own docs name: approximate counts, less reliable “with large numbers of small objects.”
One honest limit on this test: it is one model, one chart and one day, with questions we wrote ourselves, not a study across providers, chart styles or harder questions. Treat it as one concrete, reproducible data point, built from a real file with a real date on it, rather than a general verdict on any vision model. The questions and the chart-drawing method are described above in full, so the same test can be rerun on a fresh file at any time.
A structured read skips the picture. Instead of a model estimating a number from where a line sits against an axis, it works from the real OHLC (open, high, low, close) candles and the real indicator values computed from them, the same inputs any charting platform uses to draw the picture in the first place. Nothing in that path gets resized before anyone looks at it.
Done properly, a structured read covers more ground than one screenshot usually shows at once:
The one trade-off: a structured read needs the data lined up in the first place, whereas a screenshot needs nothing but a picture. Our prompt library has ready-made prompts for both approaches, including one built specifically to pull levels from a data table instead of a chart image. The same real-data-versus-guess split runs through AI backtesting: a model that computes from real prices, or one that only describes a plausible result.
Two habits carry most of the benefit: make the model check what it can actually see before it says anything, and prefer handing it numbers over handing it a picture whenever you have the choice.
Before you read this chart, tell me: can you clearly make out the price axis labels? Is the chart resized or compressed in a way that makes fine detail hard to read? For any price or level you give me after this, mark whether you read it directly off a labeled gridline or estimated it between gridlines, and say which estimates you’re least confident in. Don’t give me a single number without one of those two tags.
Here is the real OHLC data behind this chart as a table (date, open, high, low, close), not an image: [paste the rows]. Using only these numbers, give me the trend on this timeframe, the two nearest support and resistance levels with how many times price touched each one, and whether momentum is confirming or diverging from price. State your method in one line, so I can check it, and don’t describe anything you can’t trace to a row in the table.
Our wider library of ChatGPT and Claude trading prompts has more for both the screenshot and the data-table approach, including a four-timeframe read and a prompt built to pull swing highs and lows from a price table. The two above are about the habit underneath any of them: separate what the model can prove from what it’s guessing, every time.
Ethy’s chart read is the structured path by default, with no screenshot involved at any point. Ask about a market and it reads the daily, 4-hour, 1-hour and 15-minute charts together; ask about just one timeframe and the read still brings in the daily, 4-hour and 1-hour alongside it, so a 15-minute signal is never judged without the bigger picture around it. Ethy’s AI chart analysis lists everything a read covers.
Because every one of those is worked out from candle and indicator data, none of it goes through the resize-and-estimate step a screenshot takes: there is no image to compress, no gridline to eyeball, and no candle too thin to call. Ethy’s charts are drawn with TradingView’s charting library, fed with Ethy’s own market data; there is no link to a TradingView account, and Ethy doesn’t import Pine Script, read TradingView alerts, or connect to any outside assistant. If you specifically want Claude working with TradingView’s own charts and data, that setup is in how to connect Claude to TradingView with MCP.
Ethy is software for research and analysis. It doesn’t give investment advice, and a chart read, structured or not, is a read of the past, not a promise about what comes next.
Yes, it can describe the trend, call out levels and name patterns from a pasted image. OpenAI’s own vision guide warns it struggles with precise spatial localization and with graphs where colors or line styles vary, so treat any specific price it reads off the picture as an estimate, not a measurement, and check it against the axis yourself.
It can describe the shape of one well. Claude’s vision docs say its coordinate and counting outputs are approximate, and that it shrinks images larger than its limits before reading them, which is where fine chart detail gets lost. In our blind test, Claude Sonnet 5 placed the high and the low about right but overcounted red candles: it guessed 33 to 35 of 60, and the real count was 31.
Because a trend is a shape, which survives being resized and estimated, while an exact price is a single pixel position against an axis, which doesn’t. The vendors’ own docs describe this directly: precise spatial localization and fine counting are named limitations, while describing an overall pattern is not.
Data, whenever you have it. A model working from a table of real OHLC rows is reading numbers, not estimating them from pixels, so it can give you an exact support level and a touch count instead of an approximate line. A screenshot is still useful when you don’t have the data handy or want a quick visual gut check.
Its chart read doesn’t start from a picture. Ethy works from price and indicator data across the daily, 4-hour, 1-hour and 15-minute timeframes, so there’s no resizing step to lose detail in.
Treat it as a second opinion, not a replacement, on either path. A screenshot-based read carries the estimation limits above; a structured read is only as good as the data and rules behind it. Either way, Ethy presents a chart read as research, not advice, and the decision stays yours.
Trading involves risk, including the loss of your money. Ethy AI is a research and trading tool and does not give investment advice or recommendations. Backtests are simulations on historical data and do not predict future results.
Deploy your AI trading agent in one click.
