Spoiled by Fable

...like taking a Ferrari into rush-hour traffic.

Like most software engineers, back in June I had access to Claude Fable 5 for two days before it was yanked. Going back to Opus felt like a massive downgrade that I noticed in every session.

Opus is a good model. I've shipped a lot of real work with it and never had much to complain about. Two days on Fable spoiled that.

What two days bought

I spent that window on Neongrave, a shoot-em-up I'm building in Unity. It's early, so most of the work is content: enemies, bosses, power-ups, items, stages. I keep a GitHub backlog with one issue per piece, and the loop rarely changes. Grab an issue, build it, test it, playtest, then fix whatever the playtest turns up.

With Opus, I was the one catching most of the bugs, so playtesting doubled as QA. With Fable I tried something I wouldn't have risked otherwise and dropped a whole update into a single prompt: new stages, new enemies, a fresh set of weapons, a couple of bosses. It worked in coherent 30-minute stretches and handed back changes that actually held together.

The speed wasn't the part that stuck with me. It was where my attention went afterward. With fewer bugs making it to playtest, I could stop hunting for broken things and start asking the questions that matter. Is this boss fun to fight? Does this weapon feel good to fire? That's the part of building a game I actually enjoy.

The downgrade

Then Fable was gone and it was just me and Opus again. The loop still worked, but the bugs came back to playtest with me, and more changes needed a second or third pass. Nothing was broken. It was just heavier, and I only noticed how heavy because Fable had briefly taken the weight off.

Plan B on my own hardware

Losing Fable overnight made something obvious that I'd been happy to ignore: I don't own any of this. Access can vanish, get more expensive, or change under me, and I get no say in it. So during the dry spell I went back to running local models, this time on my own graphics card with LM Studio.

I tried three:

• Qwen 3.6 27B, a local code-and-reasoning model.
• Qwen 3.6 35B-A3B, fast enough to hit around 200 tokens per second on my machine.
• Gemma 4 31B, the well-rounded one, and it handles images too.

I wired a couple of them into local coding loops. They're quick, and once the GPU is paid for they cost nothing per token. They're also well behind the hosted models the moment a task gets hard. That gap is what made the detour worth it. Watching a local model take the same task and fall short showed me exactly what the frontier models had been doing for me, and "use the right model for the job" stopped being a throwaway line and turned into something I could measure.

Access again, bigger targets

When Fable came back this weekend, rationed on my Claude plan, I aimed it at more than game content.

Apsis is a space sim I'm building in Rust with Bevy, somewhere in the neighborhood of EVE Online. I asked for a lot at once: a procedurally generated network of star systems, each with its own star, planets, and moons, plus stations, a stargate network to jump between them, pirates, and enough combat to make the pirates worth taking seriously. Fable built the whole thing and, more importantly, kept all those systems moving toward the same goal without me refereeing every step. Opus could have built any one piece on its own. What set Fable apart was how little I had to coordinate between them.

Narrato keeps me honest. It's a conversational business-intelligence tool, and I'd already built the core: event ingestion, the datastore, the API. Fable filled in the console, the docs, the conversational agent and its tools, and a streaming chat interface, faster than I'd managed on my own. It's not finished. The interface needs another pass, and that pass is on me to write sharper prompts about how it should feel to use. Fable closed the gap on the plumbing quickly, but the product decisions were still mine.

How I split the work now

Fable is the best model I've used, and it's also overkill for most of what fills my day. Routing everyday work through it is like taking a Ferrari into rush-hour traffic. All that headroom, and you're still crawling next to the minivan.

So I stopped reaching for it by default and started matching the model to the task. I've found the sweet spot for Fable is architecture and planning: turning fuzzy goals into concrete issues that a less capable model like Opus can confidently build. Sonnet is my go-to for routine tasks, simple changes, and the tedious/repetitive workflows. Local models sit underneath all of it, for the days I'd rather not depend on anyone's API.

This wasn't a profound lesson; in fact, it's one that gets repeated often. But for some reason, getting access to an upgraded model like Fable, losing that access, and getting it back on a tight token budget really crystallized the concept for me: use the right model for the job. Which is most of engineering anyway, once you know the constraints.