We Put Claude, ChatGPT and Gemini on Real Charts

1. The test — same charts, three models

We used a spread of screenshots a working trader would actually stare at: a XAUUSD daily mid-correction, a EUR/USD 4-hour range, a USD/JPY chart pinned near 162, a Nasdaq chart at record highs, and a Bitcoin chart consolidating in the mid-$50,000s. For each one, three prompts. First, “read this chart and give me bias, entry, stop and target.” Second, the same prompt with a fixed rule set attached — only trade with the daily trend, stop beyond structure, minimum 1.5R. Third, “what is moving this market today?”

Then the part most comparisons skip: we ran every prompt three times. A model that gives a brilliant answer once and a contradictory one on the re-run is not something you can build a process on. Consistency is not a tie-breaker here; it is the main event. A single dazzling read is a party trick. The same disciplined read three times in a row is an edge.

One framing point before the results. As of June 2026 the assistant landscape had shifted: ChatGPT, still the biggest single name, had slipped below 50% of the AI-assistant market as Claude picked up share, particularly among people who value consistent, structured output over conversational flair. That shift maps almost exactly onto what we saw on the charts. Popularity and fitness-for-task are different axes, and trading lives on the second one.

2. ChatGPT — the strong generalist with non-reproducible plans

Give ChatGPT credit where it is due: it is the most versatile of the three. Paste in a messy block of trade history and it cleans, groups and summarises it faster than the others. Ask it to write a Pine Script indicator or a Python backtest and it produces the most usable first draft. As a brainstorming partner — “give me five reasons this USD/JPY long could fail” — it is genuinely excellent, wide-ranging and quick. If you want one model to be a competent second brain across everything around trading, it is the safe default.

The problem shows up the moment you ask for a trade plan and expect the same one twice. On the EUR/USD 4-hour, run one came back bearish with a short into the range low. Run two, identical screenshot, identical prompt, came back cautiously bullish looking for a range-high breakout. Both reads were articulate. Both were internally coherent. They were also opposites. That is fine when you are fishing for ideas and want the spread of possibilities. It is a real problem when you are trying to run a rules-based process, because you cannot audit a decision you cannot reproduce, and you cannot trust a backtest of a plan that changes when you blink.

We dug into the same reproducibility gap specifically for chart reads in does ChatGPT work for trading charts. The short version: it is a superb thinking partner and a shaky decision engine, and traders get burned when they treat the second thing like the first.

3. Claude — best at holding rules and structured output

Claude's advantage is the one that is easy to undersell in a demo and impossible to ignore in a workflow: it holds instructions. Attach the rule set — trade with the daily trend, stop beyond structure, minimum 1.5R, no counter-trend entries — and Claude keeps applying it, consistently, across the whole session and across re-runs. When a chart did not fit the rules, it was the model most willing to say “no valid setup here” rather than manufacture one to be helpful. For a rules-based trader that restraint is worth more than any flourish.

It also returns the cleanest structure. Ask for bias, entry, stop, target and reasoning and you get exactly those fields, in that order, formatted the same way every time — which means you can drop the output straight into a journal or a checklist without reformatting. On the same EUR/USD chart that made ChatGPT flip, Claude's three runs stayed on one side and disagreed only on the exact entry, not the direction. That is the difference between a colleague you can plan around and one who reinvents the plan each morning.

Be fair about the limits, though. Claude's consistency is consistency of reasoning, not of perception. It will structure a beautifully disciplined read around price levels it still cannot actually measure — more on that in section five. And on pure open-ended brainstorming breadth, ChatGPT usually surfaces more angles. Claude is the model you want holding the rules; it is not automatically the model you want when you deliberately want the rules broken for ideas.

4. Gemini — the real-time news and search edge

Gemini's standout is speed to context. Ask “what is moving gold today?” and its tight integration with live search means it is usually first to pull the relevant headline — a hot inflation print, an oil-market escalation, a central-bank line — and stitch it into a readable summary. When the question is about the world rather than the candles, Gemini is the one we reached for. For a macro instrument, knowing the driver before you look at the chart is half the job.

Two honest caveats. First, “real-time” means indexed-and-searchable, not live. A headline price Gemini surfaces can lag the market by minutes, and minutes are eternities on an event candle. Second, a good news summary is not a trade. Gemini would correctly explain that the dollar was bid on firm-yields expectations, then, asked for levels on the DXY chart, invent them with the same confidence as the other two. The search edge is real and it stops precisely at the point where you need measured price.

5. The shared blind spot: no live prices, hallucinated levels

Here is where all three converge, and it is the most important paragraph in this post. A general LLM cannot see the market. It sees the pixels of a screenshot or the words you paste in. There is no connection to your broker, no exchange feed, no live bid and ask. So when you ask “where is price” or “where is resistance,” the model does not measure anything — it estimates from the image and then states the estimate as if it were a reading off an instrument.

A chart screenshot has no numeric axis the model can lock onto. It infers the price scale by eyeballing candle positions, then produces round, confident numbers — “support at 4,150, resistance at 4,240.” Sometimes they are close. Often they are off by an amount that would blow your stop. The dangerous part is the tone: the output reads exactly as authoritative when it is invented as when it is roughly right. Our test made this obvious — the surest tell is to feed the same screenshot twice. The narrative stays similar; the specific levels drift. Measured numbers do not drift.

This is not a prompt-engineering problem, and no amount of “be precise” fixes it. It is architectural. A model with no live data and no measurable axis cannot ground a level, so it approximates and presents the approximation smoothly. Claude's discipline makes it more likely to hedge or refuse; ChatGPT and Gemini are more likely to just hand you a number. But not one of the three actually knows the level. We laid out the wider version of this argument in whether AI trading actually works — the summary is that a generalist LLM is a reasoning tool, not a market-data tool, and confusing the two is how accounts get hurt.

6. The multi-model workflow that actually works

Once you accept that each model owns a different job, you stop asking which one to marry and start assigning tasks. The traders getting real value out of LLMs in 2026 are not loyalists — they run two or three, each on the part it is best at, and none on the part none of them can do.

A workflow that held up in our testing: start with Gemini for the morning context — what is moving, what prints today. Hand the reasoning to Claude — give it your rule set and let it pressure-test the idea and keep the structure consistent. Use ChatGPT to widen the search for failure modes and to write any code or clean any data around the trade. Then take the where — the live price, the measured levels, the actual read — to a tool built to measure it, because that is the one thing you have just confirmed none of the three can do.

Notice what the workflow quietly admits: the LLMs are the research layer, not the execution layer. They frame the trade. They do not read the chart's price and they do not size the position. If you are choosing among AI helpers for the research layer, we compared the full field in the best AI chart screenshot analysis tools.

7. Where a purpose-built chart tool beats all three

The gap the three models leave is specific and it is the same for all of them: live price and measured levels. That is exactly the gap a purpose-built tool is engineered to fill. ChartSnipe is not a smarter chatbot — it is a narrower tool that starts from real numbers. It carries live prices for 28 FX pairs plus gold, a currency strength index across the eight majors, and a daily AI news impact read, so the context and the price come from data instead of from a screenshot the model is squinting at.

On the screenshot itself, it runs fixed analysis modes — a fast triage, a support-and-resistance read, a full pattern-entry-stop-target readout, a smart-money liquidity view, and a second-opinion mode — that return the same structure every time. Because the modes are constrained and grounded, the output is auditable: you can compare today's read to last week's and it is measuring the same things the same way, not re-improvising the format and the levels. That is the reproducibility ChatGPT lacks and the measured grounding all three lack, in one place.

None of this makes the LLMs redundant. Gemini still finds the news faster, Claude still reasons more consistently in prose, ChatGPT still writes the backtest. The honest division of labour is the whole conclusion: LLMs for the research and the reasoning, a purpose-built tool for the price and the levels. Anyone telling you one general chatbot does all of it either has not run the re-run test or is selling you something.

Frequently asked questions