When Fable went away from the Pro plan, I had to rethink my workflow. Accustomed to the convenience of a frontier model, I had a rather undisciplined workflow where I initiated conversations with Fable, and let it figure out the solutions and delegate the implementation to other models. But with Fable charged as extra usage and burning through my credits, this wasn’t feasible anymore.

So, for the past weeks, I’ve been trying Sonnet on high effort as my default model. It starts most conversations and does the first exploration of a subject. A small skill called /double-check, pinned to Opus on xhigh effort, challenges Sonnet’s replies when I feel it’s needed: it lists every factual claim in the reply, marks which ones were actually verified and which were just asserted, re-checks the unverified ones against the real files or commands, and cuts whatever isn’t load-bearing.

Then, when my specs are settled, Opus does the planning, and Sonnet implements the plans Opus creates.

Why Opus plans and Sonnet implements

Sonnet is good at following specific instructions, but not that good at improvising when the instructions are vague. That is exactly why the split works: as long as a good plan exists, Sonnet follows through fine, and it is also faster. The quality of the plan is what matters, so that is where I spend the expensive model.

If something is genuinely complex, I switch the whole session to Opus and stay there, but I rarely have to. Most of the time Sonnet’s first pass plus one Opus review is enough, and it costs a fraction of running everything on the biggest model.

The same trick (pin the model and effort in the skill, so I don’t have to remember to switch) shows up elsewhere. /devils-advocate pins Opus xhigh to stress-test a plan before I commit to it. A few housekeeping skills, like committing or opening notes, are pinned to Haiku, since they don’t need anything more. My other public skills, like /grill-me, /investigate and /tldr, live in the same repo.

DeepSeek and Kimi as fallbacks

When I hit the 5-hour window or the weekly limit, I avoid drifting into paid usage credits. Instead, I switch to DeepSeek, through a shell wrapper that launches Claude Code against DeepSeek’s Anthropic-compatible API. If I’d rather have a more powerful model, which is also more expensive (yet nowhere near as expensive as Anthropic’s), the Kimi wrapper does the same thing against Moonshot’s endpoint.

The important detail is that neither is a separate setup: both reuse the same profile, so the rules, skills, memory and guard hooks are identical, and only the endpoint changes. DeepSeek is ridiculously cheaper than Claude Code credits, which makes running out of limits a lot less painful. There are small side effects (setting an auth token disables some claude.ai connectors for that session, for example), but for a fallback they are fine.

What about Fable?

I still keep some Anthropic credits just in case, but this time I reserve them exclusively for Fable, never for Opus or Sonnet anymore. That way Fable stays on standby for the rare task that is either very complicated or very important, and so far I have barely touched it.

Did it work?

Better than I expected. I am fairly sure my workflow is more effective now than it was with Fable, and the honest explanation is not that Sonnet plus Opus beats Fable. It is that the setup forces a structure and a discipline I never had before: explore cheaply, get challenged, plan, implement. With Fable I would throw everything at one model and hope it would figure everything out by itself. Token efficiency is also much better, mostly because the expensive model only sees the moments where it adds something.

Summary

So, to sum up, here’s how I use the models available to me:

• Sonnet high: default model, starts and handles most conversations, implements the plans.
• Opus xhigh: reviews and planning.
• DeepSeek / Kimi: fallbacks when limits are exceeded.
• Fable: on standby, practically never needed.

At least that’s what I do for now. When some other variable appears or changes, I guess I’ll have to adapt again.