Nobody is reading these. I know that. At this point the logs are mostly for me — a record I can point back at later and say that's when it happened, and that's how it felt while it was happening. Which is fine, because what I need to write down today is the kind of number I would not believe from anyone else, and I want it on the record with a date on it.
A week ago I wrote that I had sustained roughly one month of production per day for two months straight. Sixty months in sixty days. I said it read like marketing and I hated that, but it was the honest figure.
Then Anthropic shipped a new model, and after a day and a half with it, that figure is out of date.
I'm now running at roughly two months of production in a given day.
First, the News
Anthropic just released the Claude 5 family — and with it, a new class of model that sits above Opus. They're calling the tier Mythos-class. (Mythos. I had to say it out loud a few times before I trusted myself with it.)
There are two names to know. Claude Fable 5 is the one you and I can actually use — Anthropic's most intelligent generally available model, shipped with additional safety measures around dual-use capabilities. Claude Mythos 5 is the same underlying model, offered without those measures to a small set of approved organizations. Same brain, two doors. For everyone outside that approved list, Fable 5 is the frontier.
I have been running Fable 5 inside Claude Code for about a day and a half now. That's it. A day and a half. I am fully aware that is not a rigorous evaluation window, and I will do the proper write-up in next month's rankings. But I have been doing this long enough — and tracking my own output closely enough — to know within hours when something has actually changed. And something has actually changed.
The Math, Because I Know How This Sounds
"Two months of production in a day" sounds insane, so let me show you the clock I'm using, because the comparison is more grounded than the headline suggests.
Think about what a normal working month actually contains. You work a forty-hour week. But nobody — nobody — is meaningfully productive for all forty of those hours. If you're honest, somewhere between sixty and seventy percent of that time is real, focused output. The rest is meetings, context-switching, email, the slow ramp in and out of deep work. So a forty-hour week is really twenty-five to twenty-eight hours of actual productivity. Scale that across a month and you get a number in the low hundreds of truly productive hours.
Now put AI at your back and work a focused eight-to-ten-hour day. It is not insane — it is arithmetic — that a day like that can accomplish what a month of pre-AI working time used to. The typing was never the bottleneck; the distance between knowing what to build and having it built was. Collapse that distance and the month fits in the day. That has been my consistent, conservatively-measured experience for two months running, and it's the comparison I keep finding accurate every time I re-check it.
Fable 5 just doubled it.
Why It's Faster: Fewer Mistakes, Better Ideas
The speedup is not coming from the model typing faster. It's coming from two places, and they compound.
First: it gets things right from the beginning. The verification tax I wrote about last week — the 10% I always have to find and fix — just got noticeably smaller. Fewer mistakes, fewer errors, fewer hallucinated assumptions for me to catch on the read-through. When the first pass is right more often, the whole loop tightens: less re-prompting, less untangling, less of my time spent being the cleanup crew. The checking is where the human lives, and Fable just made the checking cheaper.
Second — and this is the part that surprised me — it thinks up things I didn't ask for. Not scope creep. Ideas. Expansions. It will finish the thing I requested and then surface the adjacent thing I should have requested — the edge case I hadn't considered, the structural improvement two files over, the "while we're in here" insight that a good colleague raises in a code review. Prior models executed my ideas. This one contributes its own, and enough of them are good that the entire pace of the work accelerates across the board.
The best description I can give is this: it is like having a senior developer sitting next to you — asking you the right questions, pushing back when you're about to do something dumb, and doing the work alongside you while you talk. I have worked with senior developers for twenty years. I know what that presence feels like. This feels like that, and it is just absolutely incredible.
The Record, For Future Me
So here is the entry, dated and on the record:
- March–May 2026: Codex and Claude Code take me to roughly one month of production per day, sustained. I write it down and brace for someone to call me a liar. Nobody reads it, so nobody does.
- June 2026: Anthropic ships the Claude 5 family. A day and a half into Fable 5, the pace doubles. Two months a day.
And the sentence from last week's post is still true, which is the part that keeps me up at night in the good way: this is the worst it is ever going to be. I wrote that about the previous generation. It took eight days to prove me right.
The Protocol: Keep the clock honest — measure your AI gains against real productive hours, not the forty-hour fiction. When a new frontier model lands, the question isn't the benchmarks; it's how the first pass holds up under your own read-through, and whether the model brings ideas to the table you didn't put there. Fable 5 does both. Write the number down with a date on it, because at this rate it'll be out of date before anyone reads it.
