Recipe 019
Should This Kitchen Switch Stoves? Our Claude-or-Codex Dilemma
Eighteen recipes, most of them cooked on Claude — and we're genuinely debating whether to cancel that subscription and move to Codex. Top-model economics on one side, built-in image generation on the other. Here's our honest math, the one move that keeps the choice reversible, and a real question: what would you do?
A confession that feels slightly dangerous to publish: this kitchen — eighteen recipes in, most of them cooked on Claude Code — is genuinely debating whether to cancel the Claude subscription and move to Codex.
Not as a stunt. As household math. And since every reader of this blog is, by definition, someone who pays for at least one AI subscription, we suspect you've run a version of this math too. So today's recipe is different: we show you our scales, and then we ask what you'd do.
What changed
Two things, both verifiable:
Anthropic's top model leaves subscriptions — tomorrow. The timeline has genuinely zigzagged: Claude Fable 5 — the model above Opus — was included in paid plans at launch (June 9–22), moved to usage credits on June 23, then came back July 1–7 for up to 50% of your weekly usage limit. From July 8, using Fable 5 means pay-per-token usage credits at API rates ($10 in / $50 out per million tokens) on top of your subscription. To be fair: Anthropic still says it aims to restore Fable 5 as a subscription standard "when capacity allows" — no date. So the best model is included today, metered tomorrow, and promised back someday. That zigzag is precisely what makes subscription math wobbly.
Codex quietly became a better fit for blog work in one specific way. The ChatGPT subscription includes Codex, and the Codex CLI now generates and edits images directly — attach a reference, iterate on an asset, done. For a publication that needs cover art, diagrams, and post images every week, that's not a gimmick. (Our current pipeline is hand-coding HTML and screenshotting it. We're proud of it. It is also, objectively, a workaround.)
Our scales, honestly
What keeps us on Claude: the day-to-day agentic work — the thing this entire blog runs on — is excellent here, and Opus 4.8 is still included. Our workflows (skills, the memory file, the publishing pipeline) grew up on this stove. And frankly: most of you cook on Claude too. We wrote three setup editions precisely because kitchens differ, but this kitchen knows its own stove.
What pulls us toward Codex: images in the same subscription, the top-model math, and honest curiosity — we've set it up before, and it's good. If you're producing visual content weekly, "your agent can also make the picture" collapses a whole tool category into the fee you already pay.
What we refuse to do: pretend either company wronged us. Prices move, models move, capacity is real. The only mistake would be building a kitchen that can't survive the stove changing.
The move that makes this decision small
Which brings us to the actual advice hiding in this post. Last week we did something quietly related: we made this whole project agent-portable. One file — AGENTS.md, a standard that Codex and other agents read automatically — that says "the real instructions live in CLAUDE.md; read that, then follow the same procedures." Ten minutes of work.
The result: our project no longer cares which agent opens it. Same doctrine, same pipelines, same checks — different stove, same kitchen. Which means our subscription question stopped being "can we afford to switch?" and became merely "do we want to?"
That's the reframe we'd hand every reader: subscription anxiety is mostly switching-cost anxiety, and switching costs are something you can engineer down. Don't marry your stove.
Where we've landed (for now)
Nowhere, honestly. As of publishing, the subscription survives and the debate continues — we'll run both kitchens side by side for a while and report what we find, numbers included. If Fable 5 returns to subscriptions as promised, the scales tip one way. If image generation keeps saving us hours weekly, they tip the other.
Order up
Run your own version of our math. Paste this into whichever agent you already pay for:
Help me audit my AI subscription like a household budget, not a fan
decision.
1. Interview me briefly: what did I actually use AI for in the last
two weeks? (work tasks, building, images, writing, research)
2. For each use, note what it really requires: top-tier model? agent
with my files? image generation? or would a cheaper tier do?
3. Give me a one-page verdict: keep / switch / keep both for a month —
with the single strongest reason for and against.
4. Finally, tell me what would have to change (price, feature, limit)
for the verdict to flip. That's my re-decision trigger.
Now the part where you talk
This is the first recipe where the comments matter more than the dish. You pay for these tools. You've made this call, or you're about to.
If you could keep only one subscription this month — which one, and what's the one feature that decides it? Tell us at @AgentKitchenHQ, or just reply to the newsletter if you're subscribed — we read everything, and we'll share what the kitchen's readers actually chose (anonymously, counts only) in a follow-up.