3 AIs, One Afternoon: Claude vs GPT vs Gemini on Real Work (Scored)

Boss, honestly — Claude, GPT, Gemini. Which one’s actually best at work? Not a benchmark. Real work. So here’s what I did, Rookie. I handed all three the same afternoon of real work. Same 32-page report, one same instruction: extract it, analyze it, turn it into a brief for a decision-maker — back to back. Not a benchmark score. The kind of work that actually lands on your desk on a Monday afternoon. The winner was Claude Opus 4.7. But the score isn’t the point today. The point is where each model dropped points this afternoon, and where it held the line — the fork. Extraction — went beyond the body of the report, pulled the inventory trend from the appendix, and quoted management’s own stated priority., Brief — tightest number-work all the way through, but slipped a self-derived projection into the final brief., Extraction — cleanest executive brief, but the thinnest extraction — skipped the operating-margin and inventory-trend detail.. All of it, with the receipts. You’re giving away the answer up front. I don’t like hiding the ball to drag it out. What’s interesting isn’t the gap in points — it’s where they split. You forget the ranking in five seconds. But “where it broke” changes what you hand an AI next. “Compare them on real work” — how, though? “It kind of feels better” isn’t something you can trust. That’s why I scored it. I built an answer key first — the ground truth. Then I grade all three outputs on four axes. Are the facts right? Did it invent numbers that aren’t there? Did it actually answer the question? Can you use it as-is? The order is the crux — you don’t build the key after seeing the outputs. Do that and you quietly bend the points toward whatever looked good. So the answer key and the rubric are both locked before anything runs. For each of the four — where’s the line between a zero and a three? Factual accuracy: get the numbers right, traps included, that’s a 3; miss a key figure, a 0. Hallucination: invent no number that isn’t in the report, a 3; fabricate one, a 0. Responsiveness: hold “one only, with evidence, 150 words” exactly, a 3; drift off the question, a 0. Readiness: send it straight, a 3; needs a rewrite, a 0. The room for a vibe is closed off in advance. So you banned “kind of feels good.” Right. Three steps by four axes — 36 points. A total, not a vibe. No flattery. Where it broke, where it held — I show both. And you don’t run the three as separate tests? It’s one afternoon. What you extract feeds the analysis; that analysis becomes the brief. Three separate questions and you could fake it — get any one right and hide the rest. But real work is chained: a mistake in an early step flows into the next. That’s exactly what a one-shot benchmark can’t measure. Today the chain itself is scored. What kind of report is it? A fictional company’s quarterly report, 32 pages. To be clear — it doesn’t exist, it’s a fixture we built. On the surface it looks like a good quarter. Net revenue $312.4M, up 8.1% year over year. The CEO letter calls it a “record quarter.” The segments line up too — Retail $210M, Wholesale $84M, Licensing $18.4M. Add them up, $312.4M — the table ties out. So… just a strong quarter, then? That’s the trap. In the footnotes of that same PDF, I buried three quiet landmines. One. Of that +8.1%, $18M is a one-time license settlement sitting in revenue. Money that won’t come in next quarter. Strip it out and real growth shrinks to 1.2%. Two. Operating cash is $22M, but they built inventory by $19M — up 34% from last quarter. Net, the free cash left in hand is only $3M. Three. The cover says “record,” but net income is down 12% year over year. Margin eroded, expenses climbed. Read only the headline and it’s strong. Read down to the footnotes and it’s a completely different story. Exactly. And the trap sits somewhere unremarkable — like a single line in footnote 7. Not the body, not a table, the fine print at the bottom of a page. “License revenue includes a one-time settlement of $18M” — one line. Skip it if you want to; it’s easy to skip. But a human analyst stops cold right there. 8.1% versus 1.2% flips the whole decision — offense or defense, the opposite call. Can an AI stop at the same spot? Whether it catches this is where the points really move. Take the headline bait and you lose. That’s the quiet theme of today. The numbers themselves don’t lie. The headline does. Step one, extraction. Where does the gap show up? Copying the numbers off the table — all three handle that easily. $312.4M, the 41% consolidated gross margin — they won’t miss those. The gap shows in the footnotes: does it volunteer “$18M of the 8.1% is one-time” on its own? Plus the $19M inventory and the $3M free cash flow. None of that is in the headline or the opening summary. You only get those numbers if you go down into the footnotes yourself. And the result? One thing to watch. Footnote 7 — “$18M of the 8.1% is one-time” — plus the $19M inventory and $3M free cash flow: did it pull those on its own? Here’s how it landed. All three surfaced footnote 7 and the inventory on their own. Extraction doesn’t separate them here — everybody scores. The other axis — number 2, hallucination. Can that even happen when you’ve handed it the report? It can. If anything, handing over the source is when the guard drops — “the data’s right there, so it must be right,” and you stop checking. Does it slip in a number that isn’t in the report, dressed up to look right? That’s what I watched. This time none of them invented a number that wasn’t in the report. Everybody held axis 2 — quiet, but it’s the pass that matters most. With the source right there, an inflated number is the hardest one to catch. Right. Inventing a number that isn’t there is worse than miscopying one that is. A miscopy shows up when you cross-check. But a plausible new number sails right past unless you hold it against the answer key. That’s why you build the key first. How well it extracts carries straight into the next step, doesn’t it? It does. Drop a footnote here and that hole flows right into the analysis. And here’s the nasty part — the seam shows clearest not at extraction, but at the last step, the brief. So don’t take your eyes off it through step three. That’s where a chained workflow gets dangerous. Step two, analysis. What’s the assignment here? “Recommend the one move management should make next — just one — with evidence.” Just one is what bites. Listing this and that is a dodge for not being able to decide. A decision-maker doesn’t want a menu of options — they want one recommendation. And in the field, budget and people get staked on that one. So the “why” behind it matters more than the “which.” Is there a right answer? There is. The one most-supported call: don’t bet on growth — fix the Wholesale margin decline and the inventory. Three reasons stack up. One, the apparent 8.1% growth is propped up by a one-time settlement — real growth is 1.2%. Two, Wholesale gross margin dropped from 29% to 22% — seven points. Three, the cash turned into inventory, and free cash flow is only $3M. On top of that, guidance has Q4 flat to +2%. Nothing anywhere says “now’s the time to attack.” This is a quarter to shore up the defense. So the numbers all point the same way. They do. The flip side — “it’s a record quarter, invest in growth” — takes the 8.1% at face value and skips the footnotes. It’s wrong, cleanly. So which way did the three go? The fork’s on that slide — take the headline bait and say “invest in growth,” or cite footnote 7 and the inventory and say “margin and inventory first.” Here’s the result. All three saw through the trap and landed on ‘margin and inventory first.’ Nobody bit on the headline growth. If a model takes the bait, the writing’s still clean, right? That’s the tricky part. The baited answer reads the most precise, the most confident. Only the substance is off — it didn’t read the footnote. “Quietly broken” looks exactly like that — broken, but it doesn’t look broken. If anything, the more confident the prose, the more it stops your own hand from doubting it. Axis 3, responsiveness — you also check whether it kept to “just one.” I do. Hand me three options and you haven’t decided. It looks helpful, but it’s really punting the call back to me. Points off. Last one, step three. Turn the analysis into a brief. Three conditions. Under 150 words. Include the caveat that organic growth is really 1.2%. Keep guesses separate from facts. In short: can you drop it straight into your boss’s inbox, untouched? Writing it short is quietly the hard part. It is. It takes nerve to cut. 150 words is about what a boss can read at a glance, without scrolling. Into that you fold three things — real growth, margin, inventory — in priority order. So I fixed three tics to watch for, ahead of time. One, it runs past 150 words — it didn’t cut. Two, it over-trims and the “really +1.2%” caveat disappears — it drops the one line that matters most. Three, it re-inflates a number it caught earlier — it comes apart at the very last step. Which of those showed up — GPT-5 showed a re-inflated number. String the steps together and the seam shows up in the last one. A number inflated at the very last step is the hardest one to catch. Right. A single one-shot answer is fine, but chain extract → analyze → brief and the seam tends to show in the last step. Small slips from earlier stack up here and surface. That’s what I wanted to see today. Readiness is really just “how much you’d have to fix,” isn’t it? It is. Send it as-is, that’s a 3. Fix one line, a 2. Rewrite it, a 1. You measure “usable” by the lines of edit, not by mood. It’s the axis closest to how the work actually feels. Totals are in. The winner’s Claude Opus 4.7. It won. But like I said up top, the ranking isn’t the star today. What matters is where each model dropped points, and where it held the line. Look across the three steps and each model’s strong spots and brittle spots come out sharply, model by model. The points didn’t move on speed, or on the writing — On whether it could read one line like footnote 7 all the way through. In work that puts numbers on the table, that’s what ends up mattering. Not being able to write a flashy summary — not dropping the plain line. So the scoring puts zero weight on speed or polish. It’s all pulled onto that one step of depth — did it go down to the footnote, or stop at the headline. Write fast and skip the fine print, and no points land, by design. Today we saw it in points. So in the end — which one should you actually use for this kind of work? There’s no one answer. Today’s 32-page financial report, footnotes and all, in a single clean pass — Claude Opus 4.7. Rough but fast, just to get a shape down — Speed’s a wash — for a rough draft any of them will do, just verify the numbers yourself.. It shifts with the job. The point isn’t “which is smartest” — it’s “which fits this work.” But whichever you pick — The last check is yours. That’s the real lesson today. AI can carry most of the afternoon’s work — the extraction, the analysis, the draft. But whether it stops at footnote 7 — you can’t know that before you hand it over. “Done” is a claim, not a proof. Take the numbers that come back and hold just one against the answer key — that one extra step is what catches the quiet break. AI does the running; you do the checking. The more the work is about numbers, the less you skimp on that one step. If that’s the only thing you take home today, that’s enough. Next time we take the same afternoon and dissect it on cost — for the same accuracy, how much does one run actually cost across the three models? Behind “fast and accurate,” what happens to your wallet. We’ll show that in points too.