I don't know what the next model is for

At 22:38 on Friday 30 October I sent a launch post about opencommerce and asked whether we should list it. By 22:49 I'd sent four more (Microsoft's Mage, a DeepSeek build from Unsloth, Nace's Drex Max and Xiaomi's OCR model), and while I kept going each one was checked, and the three new models were researched and graded by an agent of their own. Two of them claimed "#1" or "first" on their own pages and turned out to be wrong on the leaderboards they named. By midnight they were listed, with the reasons and sources for every score.

That's when it started to feel surreal, with everything moving at once.

What one day was

Quynh and I build Anchor Terminal. This is Friday from git and GitHub, counted at 23:45 ^[1].

The features were the bigger part. Accounts with a folder per person and sign-ins that last while you visit. People linking their own page and writing their own about. Live stats in memory and a view counter on every page. A security sweep. The fix list and the company dashboard behind three plans priced as Linux file modes (£4.44, £6.44 and £777), free while we're in beta. Every listing's dashboard numbers worked out ahead for 60 days. Four blog posts.

I think a team of five researchers and five developers would need about two working weeks for that. A careful person grading a tool the way our checklist asks takes about four hours, so 24 listings is about 100 hours, and the features are about 35 developer days with tests and written decisions. That's my estimate and not a measurement, and it leaves out hiring the ten people, which is months.

Friday wasn't unusual. Anchor Terminal's first commit was on 25 September 2026, and five weeks later the site has answered 2,004,331 page views (people, agents and crawlers, every request answered with a 2xx or 3xx) ^[2] and lists 1,375 tools with the sources for what it says about each. Before Opus 5.5 I think it would have taken a funded team, and a lot longer than five weeks.

Why it feels surreal

I've been building software for most of my life, and I spent twelve years building CryptoCompare. I know roughly how long things take, and that sense is now wrong by a factor I can't hold in my head. Every listing still gets read, checked and argued with, every feature still gets tested, and I still read what goes out. What changed is that the waiting is gone. I ask, and by the time I've read the answer to the last thing, the next one is done.

The other surreal part is that it already does anything I need. I keep asking myself what I'd want from the next model, and I don't have a good answer. Faster, cheaper, fewer mistakes, sure. But I can't name a task in my week it can't already help with, and the next model is coming anyway, and it'll make Friday look slow.

Every model drop is an expansion

I don't know if anyone can relate to this, but it feels like going from vanilla WoW to The Burning Crusade. The gear that dropped at level 62 was better than the gear people had spent months farming. That's how every model drop feels now. A huge power spike in the first few days that puts you level with people who have been grinding the previous model for months.

It's weird. If you start something now, you have until the next model to have a bit of an edge, and when the next wave comes you're immediately behind. So I think you need to save the things you want to achieve and the methods you think are best, written down, so you can run them again on the next model on its first day. That's why every decision on Anchor Terminal goes into a written spec with the reason we took it, and every listing is graded against a checklist anyone can read. The model will change under us, and the checklist and the specs are what we carry over.

It isn't only software

Software is where this showed up first, because the work is text and the results can be checked by running them. I don't think it stays there. Picture the same Friday for a solicitor reviewing a stack of contracts, an accountant closing a month for thirty clients, a researcher reading every paper in a field before lunch, a newsroom checking every claim in a story against the source. The shape is the same. Most of the hours were reading, checking and writing things down, and that's the part that just got compressed.

What I think it does

Some of it worries me, and I'd rather say so.

Checking becomes the job. On Friday the fastest things were the agents and the slowest thing was me, reading what they made and deciding. When we got lax (and I said that night that we were starting to be), the fix was more checking. A second agent that hadn't seen the code reviewed every change to sign-ins, payments and the numbers, and the third review found a serious bug before it shipped. I think every field that speeds up like this ends up with review as the bottleneck.

Claims will outrun checks. Two of the three models we listed that night overstated their own ranking. When launches come this fast, the people making the claims have the same tools we do, and I expect more of them, not fewer.

The gap gets wider. If ten people's two weeks fits in a day for a team of two, then the teams that work this way and the ones that don't stop being comparable. I don't know yet what the ten people do instead, and I don't think anyone does.

It goes stale faster. A company we listed moved its domain after we graded it, and a benchmark index changed version under us. If the world moves at this speed, a grade from last month is history, and re-checking stops being maintenance and becomes most of the work.

What's next

I'm going to keep counting, and I'll post the next Friday's numbers next to this one. If the next model is as much faster as this one was, I'd like to know what it was for.

References

Read on 31 October 2026.

1. [1]Anchor Terminal's pull requests and main branch, counted on 30 October 2026 at 23:45 London time: pull requests created that day (#324 to #351), merged or open, commits on main between 00:00 and 24:00 UTC, and listings whose listed date is 30 October. The new listings are on the directory. anchorterminal.com
2. [2]Anchor Terminal's own request records, as the admin page counted them on 31 October 2026: page views of every kind answered 2xx or 3xx since the site started. The first commit is dated 25 September 2026. What the records keep is on the privacy page. anchorterminal.com