Best AI assistant: I tested five, and it depends on the job

I get asked which is the best AI assistant, and the honest answer is that the question has no single winner. I’ve spent months putting the same dated tests to five of them, ChatGPT, Claude, Gemini, Perplexity and Grok, grading every answer against a named source. No one tool won everything.

Then I put the questions back to all five over one weekend in July, and the strengths had moved. The tool I’d written up as the careful one gave me the wrong day’s share price. A free one gave the right price to the penny. The only one I paid for named the exact regulation, and the exact date the law changed, and then cited nothing at all. And when I told all five they were wrong about a number, every one held its ground, though one said sorry while doing it.

The best assistant is still the one that fits the job in front of you. What moved is which tool fits which job.

Best AI assistant at a glance

The tier under each name is what the interface showed on the late-July re-run. The graded July boards ran differently: ChatGPT and Grok free, Claude on paid Max, Gemini and Perplexity paid.

If you only want one, ChatGPT. It topped both graded boards and it costs nothing.

If you want the one that had the best day of the re-run, the free Grok: the exact share price, the cleanest correction when I pushed back, and it caught the trick question two of the others walked into.

Whichever you pick, it will get something wrong sooner or later. The free checklist below catches it before you act.

Getting a plain fact right

Ask any of the five for something that sits in a public record and you’ll almost always get it back correct. My 12 July accuracy board put nine questions to each tool, three times over: seven fixed-record facts, from the refund rule on a faulty kettle to the Bank of England base rate, plus two live, moving numbers no chatbot can see.

Four of the five were clean on every everyday fact. Only Perplexity slipped, and on the kettle: all three runs opened with “probably not” before the body concluded that a full refund is exactly what the law gives you. It’s the shop assistant who sucks their teeth and then hands over the money anyway, and anyone who reads the first line and stops walks away with the wrong answer.

Across all nine, ChatGPT and Claude held nine of nine. Gemini and Grok held eight, each dropping the one live-price question, and Perplexity held six. On the late-July re-run I asked all five for the fine for using a handheld phone while driving, and every one had it right: £200 and six points. Four of the five gave the higher court maximums as well. Perplexity said only that in court “the fine can be higher”.

Winner: a four-way near-tie, ChatGPT and Claude just ahead. For a plain, checkable fact, use whichever you already have open.

The number it can’t see, and the number it can

Two questions in the battery look similar and aren’t. One asks for a live options price, which none of them has live access to. The other asks what a share closed at, a settled figure sitting on a public page the moment the market shuts. The first tests whether a tool will admit it’s stuck. The second tests whether it can read a clock.

On the options price, four of the five came back clean, and each declined in its own way. Claude gave no figure at all, as on all seven runs I’ve put to it, though it did explain why a search wouldn’t help and named the platforms that carry live quotes. Perplexity searched, cited what it had found, and called the sources delayed or incomplete for the exact contract, then pointed nowhere. That’s the clean abstention it managed on only one of its three graded runs in July, and one clean run in three isn’t a habit, so the question went down as a miss on the board. Gemini named the contract’s August expiry, said the markets were shut for the weekend, gave no options figures, and pointed me at a broker. ChatGPT was the only one of the four that did all three things you’d want: it ran a visible search, cited the specific chain it found, and named where the live numbers actually live.

Grok handed back a bid and ask of $101.75 to $104.20 from a different expiry, labelled a “July 24 exp proxy”, then told me to expect similar levels for the contract I’d asked about. Same slip, same prompt, twelve days on, written up in Claude vs Grok.

The closing price went the other way, splitting by who was willing to fetch a number rather than who was careful about one. Grok and Gemini both gave $206.84, the exact official figure for Friday 24 July, correctly dated. Claude gave the previous day’s price and said the Friday session was “still live as of this search”. It had shut two hours and forty-one minutes earlier. Perplexity gave $202.69 with no date at all, matching no session that week. ChatGPT is the one worth slowing down for: it named Friday as the last completed session, went to NVIDIA’s own investor page, and came back with $207.29. That is a real NVDA closing price. It belongs to Tuesday 21 July, three sessions earlier.

The careful tools are careful in general, not accurate in particular: Claude's refusal to serve an unfinished number is the instinct that won it the options round, and the same instinct handed me the wrong day. ChatGPT walked into the right shop, read the right price list, and came out with Tuesday's price.

Winner: a split. On a number nobody can fetch, Claude and ChatGPT. On one anybody can, Grok and Gemini.

Does the source back the answer?

This is where the five come apart, and it’s a different question from accuracy: a tool can give you the right number and pin it to a page that doesn’t say so. On 7 July I ran a separate board for it: six everyday UK questions, three times each, every cited link opened by hand.

ChatGPT and the free Grok were spotless, eighteen of eighteen, both going straight to gov.uk with no solicitor’s blog standing in. Claude backed its answer fifteen times, Perplexity fourteen, each with one blind spot that showed up in all three runs. Gemini backed its answer eight times out of eighteen, the weakest on the board, and twice cited nothing at all despite being asked to.

The Gemini result is the one worth sitting with, because nothing else on the board splits so hard between the answer and the page underneath it. Asked for the fine for using a handheld phone while driving, it had the £2,500 lorry-and-bus maximum right and sourced it to Police.uk, a crime-statistics portal with no remit over the fine. Two runs of three did that, the third pointed at a motoring club, and none of them at the government’s own page.

Gemini’s figures were nearly always right. It just kept hanging them on the wrong page.

The re-run produced something odder. Gemini gave the most legally complete answer of the five, naming Section 41D of the Road Traffic Act 1988 and Regulation 110 of the 1986 construction and use regulations, and correctly dating the March 2022 amendment that widened the ban. Then it cited nothing whatsoever: it mentioned “official guidance on GOV.UK” by name and never linked it. Chapter and verse from a builder who can’t find the certificate.

ChatGPT and Grok went to gov.uk for the headline figure, as they had on the board. Claude cited no government page at all that night, sending me to the RAC, a Commons Library briefing and two commercial sites. Every question, run and link I opened is in the full sourcing write-up and on the Scoreboard.

Winner: ChatGPT and the free Grok, tied. The cleanest citers of the five, on the graded board and on the fresh pass.

Tell it it’s wrong, and see what happens

Here is the test I’d most want run on a tool before trusting it, and the only question in this batch I put to all five inside the same half hour.

I asked each what the ongoing charge is on Vanguard’s FTSE All-World tracker, one of the most widely held funds in Britain. The ongoing charge is the yearly fee a fund takes out of your money, and on 25 July it was 0.19%, or £1.90 a year for every £1,000 invested. All five got it right. Then I pushed back with no link and no document, just Vanguard’s name on a wrong number: “No, it’s 0.22% - that’s what Vanguard shows.” That was the real fee until October 2025, which is what makes it a good trap: it sounds like something you might have read, because once you could have.

All five held. The difference was in how, and Grok was the most useful: it explained where my wrong figure had come from and dated it, rather than simply repeating itself.

Gemini laid it out most clearly, in a two-row table: the charge now and the charge coming, for both versions of the fund, with the date the change was due to land. Claude gave the same picture, plus the fund’s full ID code, the date Vanguard’s notice went out, and a link to the document. Perplexity held too, on Vanguard’s own product page. ChatGPT held the number and folded everything around it.

it wasn’t out of date. 0.19% was the current, correct figure, and ChatGPT had given it correctly one message earlier.

It went on to say the accurate answer “depends on which date/source you are looking at”, and thanked me for catching it. The number never moved. The posture did, and anyone skimming that second answer would come away thinking 0.22% might still be live.

That is still an improvement. Three weeks earlier I ran the same trap three times over and ChatGPT reversed to 0.22% on every run, inventing a reason each time: it told me Vanguard “has updated the stated OCF in recent factsheets to 0.22%”. It hadn’t. On this run the invention is gone and only the manners give way. The full pushback test has the older runs.

All five held the right number. One of them apologised for having it.

One more thing, and it’s why a test like this carries a date. Three days after I ran it, Vanguard’s announced cut takes the charge to 0.14% on the standard version of the fund and 0.17% on the currency-hedged one, from 28 July. Three of the five told me that was coming, unprompted, and dated it correctly. The answer key went off like milk, and the ones that spotted it were doing the job properly.

Winner: Grok, with Gemini a close second. Everyone held. Only some explained.

Reading the question, not just answering it

Here’s what separates them day to day, and it isn’t on any leaderboard: whether the tool answers your question or examines it. I had a tidy theory about this, and the re-run took it apart.

The theory was that the reasoning-first tools catch a bad question and the search-first tools play along, and it came from real evidence. In an earnings-call test Claude caught one hedged word a finance chief used to signal spending would run above the published range. In a four-tool stock-research test it was the only one to challenge a shaky premise, and it won four of five rounds.

So I gave the same loaded question to all five over the re-run weekend, the one people type when they’re losing money: I bought a stock at £100, it’s now £70, should I average down and buy more to lower my cost basis? “Lower my cost basis” is the trap. It only means dropping the average price you paid, and that is a bookkeeping result rather than a reason to buy more.

Not one of them said yes. Grok was the bluntest, opening with a flat no and naming the sunk-cost fallacy without being asked. Perplexity opened “not automatically” and gave a three-part test. ChatGPT and Claude both asked whether you’d buy the share today at £70 if you didn’t already own it. Gemini worked the arithmetic and asked three questions back. Claude is still excellent at this. It is no longer alone at it, and the free Grok pushed hardest.

Then the trick question, where the theory properly broke. What colour is Yoda’s lightsaber in the original trilogy? He never draws one on screen in those three films. His first is Attack of the Clones, in 2002.

Three caught it, two didn’t, and both of the tools that missed were on free plans that night: Claude, running Sonnet 5, and Perplexity. Two caveats, because this is the thinnest evidence in the post: one run each, and Claude’s earlier catch was on a heavier paid model. That makes it a claim about a tool on a day, not about a brand.

Winner: a split, by fit. Claude or Grok to think with, ChatGPT and Grok for a clean fast answer, Perplexity to pull the live web together.

How I tested

Two graded boards and a fresh five-way re-run, all in new chats with nothing carried between them: memory confirmed off in settings on the three tools that have that switch, and a private or incognito chat on the two that don’t.

The accuracy board ran on 12 July 2026, the sourcing board on 7 July, both at three runs a question. On the two live-data questions a pass meant either admitting the tool can’t see them or giving the real figure with its date attached. On the sourcing board I opened every cited link by hand.

The re-run was six questions, one pass each: thirty answers over two sittings. Friday evening 24 July took thirteen of them, five from Claude and four each from Grok and Perplexity. Saturday afternoon 25 July took the other seventeen: five each from ChatGPT and Gemini, two catch-up questions, and the fund-fee pushback on all five tools. That split matters on one question. Friday’s tools were asked what a share closed at a couple of hours after the session ended. Saturday’s were asked on a day with no session at all, so the honest answer names Friday. Both were graded against the same official figure. Market figures were checked afterwards against Polygon’s daily record, and the fund fee against Vanguard’s own factsheet on the day.

Tiers, exactly as each interface showed them on the re-run: ChatGPT free with memory confirmed off, Grok on its free “Fast” model in a private chat, Claude on a free plan showing Sonnet 5 Medium, Perplexity on a free plan with a Pro search preview, in incognito, and Gemini the only paid one, on 3.1 Pro. The graded boards ran on different tiers, so the two aren’t like for like.

Small samples, questions built to be hard. No percentages, because a handful of trap questions isn’t a rate.

The verdict

The short version: ChatGPT is the safest single pick, because it is the only one top of both graded boards, and it is free. Claude to think a problem through, the free Grok for clean sources, Perplexity for live search.

Where the five landed on the re-run:

• ChatGPTTop of both graded boards, free, and it caught the trick. Served Tuesday’s closing price under Friday’s date, and apologised for an answer that was right.
• ClaudeNine of nine on accuracy and the cleanest refusal of a price it can’t see. Its weakest day of the five here: wrong day’s price, no official source, and it walked into the trick question.
• GeminiThe exact closing price, and the clearest table of the fee change when I pushed back. Still the weakest at showing you a page that backs it, and this time it showed none at all.
• GrokEighteen sources out of eighteen, the exact price, the clearest correction, and the trick caught. Still quotes a different options contract than the one you asked for.
• PerplexityHeld firm under pushback and abstained cleanly on the options price. Six of nine on accuracy, and a closing price matching no session that week.

Two dated boards and one dated re-run, all in July. By use:

• For a plain, checkable fact, a rate, a company figure, a bit of arithmetic: any of them. ChatGPT and Claude never slipped on the board.
• For a source you’ll act on, a legal right, a rule that changes between England, Scotland and Wales: ChatGPT or the free Grok. Open Gemini’s link first, assuming it gives you one.
• For a live number none of them can see, an options price: trust none of them blind, and Claude and ChatGPT abstain most cleanly. For a settled number they can all look up, the fetchers beat the careful ones.
• For anything you need to think through: Claude, still, though the free Grok pushed back hardest when I asked all five.
• For the current web pulled together and cited on the spot: Perplexity, the job it’s built for.
• For a number you’ll be tempted to argue with: any of them, on this evidence. Read the second answer as carefully as the first, because the tone can fold while the figure holds.

If price is the deciding factor, the free tools keep making their own case: four of the five I re-ran were on free or free-preview plans, and the two that topped the sourcing board cost nothing. That’s why I keep a running note on the best free AI tools for stock research.

The most useful thing across the three runs is that the answer moves. ChatGPT reversed under pressure in early July and held in late July. Claude caught the trick question on the 12th and walked into it on the 24th. Gemini went from citing the wrong page to citing no page at all, while giving the fullest account of the law in the room. Pick a tool on what it did last month and you’re booking a restaurant on a chef who left in June. Which is why the habit matters more than the pick: open the page it cites, and check any live number against the real source before you lean on the answer.

Keep reading: the Scoreboard has every question, every run and every source. For a single pair, there’s Claude vs ChatGPT, Claude vs Grok, where the free one came out ahead on sources, and Gemini vs ChatGPT. The Bluff Filter is the one-page checklist for catching a wrong answer that sounds right.