We Ran the Major AI Tools for a Month: The 2026 Scorecard

Until this week, Claude Code was our main tool. It still might be next week. But somewhere in the last month — bouncing between a phone, a desktop app and three different providers to ship the same client work — the ground moved, and we noticed we'd stopped caring which logo was on the tool doing the job.

That is the whole story of AI in the middle of 2026, and most of the coverage misses it. Every week brings a headline: Grok just beat Claude, Cursor got bought, GPT-5.6 is coming, the bubble's about to pop. Read as a horse race, it's exhausting and useless. Read correctly, it says one simple thing: the stack is fluid now, and the operators who win are the ones who can move across it without flinching.

We build websites, Google Ads and SEO for clients on these tools every single day, so we ran the whole field for a month across our own setup and kept score. This is the honest version — what earned its place, what didn't, the dated numbers, and the one rule underneath all of it.

What we actually ran

The thing that makes this scorecard worth anything is that we didn't test the tools in isolation. We ran them the way you'd actually work: the same real jobs — a location-page build, an ad-account audit, a dashboard fix — pushed through different tools, with our files sitting in a provider-neutral spot so any tool could pick them up and any output could be checked against the others.

That setup is the point, not a footnote. When your work isn't trapped inside one vendor's box, "which tool is best this month?" stops being a scary, irreversible bet and becomes a Tuesday-afternoon experiment. You try the new thing on a real task, compare it to the last thing, and keep whatever won — no migration, no lock-in, no drama. We didn't run everything this round — Google's Gemini and GitHub Copilot sat it out — but we ran the tools we actually reach for. Everything below comes out of running them that way. For the record, we run these on subscriptions rather than metered API (our exact stack is below), and we haven't stress-tested any of them to the limit — so we can tell you what each tool is good at, not what it costs you per job.

Claude Code finally left the terminal

For most of its life, Claude Code lived in a terminal, which quietly kept it a developer's tool. This month we ran it properly inside the desktop app for the first time, and it changes the feel of the thing — approval cards, a background-tasks pane, work you can watch instead of squinting at a wall of terminal text.

The standout was ultracode. It's a mode that takes a single instruction and, instead of one agent grinding through it, writes an orchestration script that fans the work out across many sub-agents at once and pairs it with maximum reasoning effort — in plain terms, it splits a big job into pieces and works them in parallel. On a hard, sprawling task — the kind where you'd normally sit and babysit — it went deeper than we expected and came back with something we'd have taken a day to assemble by hand. It is not the mode you reach for to make one tiny tweak. It is the mode you reach for when the problem is big and fuzzy and you want thoroughness over speed. For that, it earned its keep.

Worth saying plainly, since we've written a lot about Anthropic: Opus 4.8 in the desktop app impressed us this month, but the top of Anthropic's own range is now Fable 5, when it's actually available (its on-again, off-again saga we've tracked separately). Our single favourite moment of the whole month was watching Fable 5 drive an ultracode run — one instruction, a dozen agents fanning out across the problem at once, the task tree filling in as we watched. It is the closest thing yet to having a team on tap. All of which matters for the scorecard later, because "is Claude still the best?" depends on which Claude you mean — and whether it's switched on that week.

AI-generated

Cowork: ambitious, and not yet for us

Cowork is Anthropic's swing at the same idea for non-coders — hand it a task and it works across your files, email, calendar and the web in its own environment. We wanted to love it, and we came in with the wind at our backs: our Claude plan had handed us double the usage for a stretch, and the idea was to pour all of that extra allowance into running Fable 5 through Cowork and get twice the work done.

It didn't pan out. Cowork was slow and buggy enough that we could barely spend the credits at all — we finished the period with Fable allowance still to burn, which is not a sentence we expected to write. The concept is right and the ambition is huge; the execution just isn't there yet. We'll check back in a few months — first swings at hard products tend to get better fast — but for now it didn't make the daily rotation.

Codex, and the wait for the next drop

OpenAI's Codex has quietly become very good, and it's the tool that most changed our mind this year. For long, grinding coding runs — the kind of big, boring job that used to eat a developer's afternoon, an hour of work spread across forty files — we've come to prefer GPT-5.5 over Opus 4.8. It holds a plan over a long horizon without wandering, and it's cheaper per unit of work when the task is mostly stamina.

We're also waiting, a little impatiently, for the next drop. GPT-5.6 — in three tiers, Sol, Terra and Luna — is already in limited preview rather than merely rumoured, with wider access promised soon. If it extends the long-horizon strength we already like, it moves up the board. The Codex app itself — desktop, CLI and cloud — is now a serious daily driver, not a novelty.

AI-generated

Cursor pulled us back

We used Cursor in our very early days, back in January 2025, and then drifted away. This month we came back, and it's the tool that most surprised us.

Two things did it. The new Cursor mobile app is a real tool, not a toy — you can kick off and steer coding agents, get notified, and merge a pull request from your phone while you're away from the desk. And its cloud agents spin up an isolated workspace in the cloud, onboard themselves to a repo, and hand back merge-ready work — no local setup, triggerable from the web, Slack or your phone. For a small team that's often out on site, that combination has quietly become our default for anything that can run while we're not watching.

The bigger picture explains the momentum. In June, SpaceX agreed to buy Cursor's maker, Anysphere, for about US$60 billion — a deal still pending a Q3 close as we wrote, that lands Cursor in Musk's orbit (SpaceX and xAI reportedly merged earlier in 2026). Put it next to xAI's Colossus buildout in Memphis, one of the largest AI training sites on earth, and the strategy is plain: own the compute, own a fast-growing tool that runs on it, and build for scale. Chess, not checkers. For you, that's less gossip than warning — the tool you lean on can be bought by a compute giant overnight, and a company that owns both the tool and the data centres under it is one step from hosting your files too. Which is exactly why you keep your files somewhere you control. (That last step is our read, not their announcement.)

AI-generated

Grok quietly closed the gap

For a long time Grok was the one you could safely ignore for real work. Not any more. Grok 4.5 landed on 8 July, and on agentic coding it is right in the mix with the frontier. The numbers xAI and early coverage published put it around 83% on Terminal-Bench-style agentic-coding tests — clear of Opus 4.8's 78.9% — though Grok isn't yet on the public Terminal-Bench 2.1 board for us to check it ourselves.

The usual caution applies, and we'll apply it harder here because we haven't reproduced these ourselves: vendor-published benchmark numbers flatter the vendor, they're a dated snapshot, and one good release is not a track record. But the direction is unmistakable — the gap that used to make the choice easy has closed to the point where, on a lot of everyday work, you often couldn't tell which model wrote the code. We'd half-suspected a strong Grok showing was coming, given the compute xAI is sitting on; the surprise was how little there now is to separate the field.

Perplexity nailed the phone

If Cursor won our desktop back, Perplexity won the commute. For everyday mobile use — a question, a document to read, a quick bit of research between jobs — it is the most polished mobile app we've used. You can upload your own files and have it work over them, and in our use it handles voice and camera input smoothly, with context carrying from phone to desktop.

What sold us was the polish and the direction. The UI is the best in the category; the "Computer" — Perplexity's multi-step agent — is a genuinely clever, faintly cute way to hand a task off; and per-job multi-model support is on the way, which is exactly how we like to work. Connecting a git repo was painless. Wiring up our own documents and secrets so Perplexity could run against them took a bit of fiddling, but we got there — and that small bit of setup is the whole "bring your own files" promise in miniature: a little cost up front, and then your stuff stays yours.

Here's the uncomfortable observation for the two big labs: for the average person doing everyday work on a phone, Perplexity and Cursor between them are already doing more than Claude and Codex. Not because the underlying models are better — often they aren't — but because the product meets people where they actually are, files and all. And between those two, the mobile apps feel very similar; what tips it Cursor's way — if the Grok 4.5 numbers aren't lying — is the model and the compute now sitting behind it.

AI-generated

The real divide: your files, or theirs

That last point is the one we keep coming back to, and it's bigger than any single tool. There's a fork in how these companies are building, and it will matter more than this month's leaderboard.

One camp is building around your files. Cursor and Perplexity both, in their different ways, assume your stuff lives somewhere you control and their job is to be the best tool for working over it — bring your own repo, your own documents, your own context. The other camp is nudging you to keep your work inside their product — artifacts, canvases and workspaces that are lovely to use and quietly hard to leave. We understand the appeal for the vendor. But from where we sit, a tool you can walk away from is worth more than a slightly better one you can't, and the lock-in camp is on the weaker long-term trajectory. Convenience that costs you your exit is not a bargain.

This isn't loyalty to a brand; it's the opposite. The reason we can be so relaxed about which tool tops the board this month is that our files never live inside any of them.

AI-generated

Where they actually land (the scorecard)

So, the honest tally. Two caveats first: the benchmark figures below are vendor- and press-reported and move constantly — treat them as a weather report, not a constitution — and the "best tool" depends entirely on the job. With that said, here's how the month shook out, and the agentic-coding numbers that were current as we wrote.

The headline from the bench: on agentic coding the field has bunched up at the top. On Terminal-Bench 2.1, GPT-5.5, Grok 4.5 and Fable 5 all sit around 83% — with Opus 4.8 a few points back at 78.9%, and Grok's figure vendor-reported since it isn't on the public board yet. A year ago the gaps were wide enough to make the choice for you. They aren't now — which is exactly why how you use the tools matters more than which one you pick.

What we'd tell you on Monday

If you run a business on these tools, or you're about to, here's the whole thing on one page.

Don't standardise on a winner. Standardise on portability. The single most valuable decision we've made is keeping our files and context provider-neutral, so trying a new tool costs an afternoon, not a migration. Do that first and every other decision gets cheaper.

Route by job, not by loyalty. Long coding grind, Codex. Big fuzzy problem, Claude Code with ultracode. Something to run while you're on site, Cursor's cloud agents. Research on your phone, Perplexity. The skill in 2026 isn't having the best model — everyone has a good one now — it's knowing which job goes where.

Judge by your outcome, not the leaderboard. A benchmark can't tell you whether a tool shipped the client site faster or saved your week. Pick one real outcome and let it decide what stays.

Stay ready to move. The tool that's ahead today won't be in three months. Build, and buy, so that's fine.

And a note that points at our next piece: the same week Grok closed the gap, Anthropic launched Claude Corps, a US$150 million program embedding a thousand fellows in nonprofits. In a field where the money mostly chases the money, teaching the charity sector to leapfrog with these tools is a different kind of bet — we'll dig into what it means for social-impact work in the next Frontier Watch.

We've mapped the money side of this shift in The 2026 Model Reckoning, the geopolitics in Who Really Controls Frontier AI?, and Anthropic's own compute bet in the Opus 4.8 breakdown. The through-line is the same: the frontier keeps moving, and the operators who thrive are the ones built to move with it. If you'd rather have a team that has already made these calls build it with you, that's what we do — come and talk to us.