Anthropic
Claude
Anthropic shipped Fable 5, and the timeline did what the timeline always does on launch day — half of it declared the singularity, the other half declared it a nothingburger. Both groups had clearly used it for about ten minutes.
I gave it a day for the launch noise to die down. Then I put it on actual work — the unglamorous stuff TerraByte runs every day. Real client builds, real automations, real debugging on code I already understand well enough to catch it lying. Two days of that is worth more than a hundred benchmark charts.
So here’s my read. Not a fanboy review, not a hater review. A practitioner’s take from someone who pays for the thing and ships with it.
This is opinion, dated today. I’m not going to quote benchmark numbers or spec sheets I can’t personally vouch for, because the launch-day numbers are mostly noise and you can read those anywhere. What I can tell you is what changed when I sat down and used it for real.
What feels genuinely better
A few things are not hype. I noticed them within the first day, and they held up.
It holds a long thread without drifting. The thing I feel most is stamina. On a big task — a multi-file refactor, a long debugging session where the context fills up — older models would slowly forget the plan. You’d watch them re-solve a problem they solved twenty minutes ago, or quietly contradict a decision you made together earlier. Fable 5 stays on the rails longer. It remembers the constraint you set at the top of the session and is still respecting it at the bottom. For agent work, that’s the whole game.
It argues with me more, and it’s usually right to. This one surprised me. I’ll propose an approach, and instead of cheerfully building the wrong thing I asked for, it’ll push back — “that’ll break when X, do you want Y instead.” A model that tells you you’re about to step on a rake is worth more than one that hands you a beautifully formatted rake. The reflexive agreeableness is dialed down, and the work is better for it.
Less hand-holding on the unglamorous last mile. Error handling, the edge case I didn’t mention, the thing that happens when the input is empty — it reaches for those without being nagged. Not always. But more often, and that’s where most of the real cost of building lives.
What’s overhyped
Now the other side, because the launch posts oversold a few things.
It is not a senior engineer in a box. It is a phenomenally fast, occasionally brilliant assistant that still needs someone who knows what “done” means. It will write code that runs and is subtly wrong. It will hand you something confident and clean that does the wrong thing in a way you only catch because you’ve seen that mistake before. If you can’t review the output, the speed is a liability, not a feature — you’re just generating problems faster. The model got better. The job of judging it did not go away.
“You won’t need to prompt carefully anymore” is nonsense. The better the model, the more your framing matters, because it’ll faithfully run with whatever assumption you left vague. Sloppy ask in, confident wrong thing out — just faster and more convincingly than before.
The benchmark jumps don’t map cleanly to my work. A few points on some eval does not translate to a few points of “my Tuesday got better.” Some days the difference is obvious. Some tasks feel identical to the last version. The headline number is real and also not the thing you’ll feel.
Where it surprised me — good and bad
Good: it’s better at saying “I don’t know” or “I’d need to see X.” That’s a real maturity gain. A confident wrong answer is the most expensive output an AI can give you, because you act on it. More of those got replaced with an honest “here’s what I’d need to be sure,” and that quietly saved me time I didn’t have to spend cleaning up.
Bad: when it’s wrong now, it’s wrong with more polish. The mistakes are harder to spot precisely because everything around them got sharper. The reasoning reads cleaner, the code looks more finished, so the one bad assumption buried in the middle is easier to skim past. I’ve had to stay more disciplined about review, not less — which is the opposite of what the marketing implies.
A better model doesn’t lower the bar for the person driving it. It raises it. The faster the assistant, the more it matters that someone in the room knows what right looks like.
And a smaller note: the obsessive engineering Anthropic puts into these tools keeps showing through. If you ever read what spilled out when Claude Code’s source code leaked — the security hardening, the startup-time tuning, the system-prompt discipline — you get why the model behaves like there’s a lot of careful work underneath it. Fable 5 feels like more of that, not less.
How I’m actually using it day-to-day
Here’s the part the reviews skip — what it does in an actual business, not a demo.
- First-draft anything. Boilerplate, the skeleton of a feature, the config file I’ve written forty times. It produces a strong first pass and I edit from there. The blank page is dead.
- Rubber-duck debugging that talks back. I paste in the broken thing and the error, and it reasons through it with me. Half the time the value isn’t the fix — it’s that explaining the problem out loud surfaces the answer.
- The work behind the work. Reshaping data, writing the glue between two systems that almost talk to each other, drafting the copy I’ll rewrite anyway. The connective tissue of a real project.
- A second set of eyes before I ship. “What breaks here? What did I not handle?” It’s good at finding the edge case I was too close to see.
What I do not do is let it touch anything a client depends on without me reading every line. The model is the apprentice. I’m still the one who signs off. That line hasn’t moved an inch, and Fable 5 didn’t move it — it just made the apprentice faster.
My read
Fable 5 is a real step up, and the step is mostly in the boring, valuable directions: stamina, honesty, judgment. Those don’t make a flashy launch graphic, but they’re exactly what matters when you’re using this thing eight hours a day on work people pay for.
It is not magic, and the people telling you it replaces the person driving it are selling something. The better the tool gets, the more the bottleneck becomes the human who knows what good looks like — which is the most reassuring possible news if that’s the job you actually do.
I’ll keep using it. I’ll keep reviewing every line. Both of those are still true, and I suspect they’ll stay true through Fable 6.
This is how we think about every tool we put into client work — capability is the easy part, judgment is the job. If you want AI doing real work in your business with someone accountable for the output, that’s the conversation worth having. See how we build AI agents that earn their keep →