Why bring your own engine
Most AI products resell you tokens with a margin on top. That model punishes exactly the behaviour you want from an agent platform — running often, retrying when something fails, checking its own work — because every one of those is billed to you twice.
Honey Badgers does the opposite. You connect the engine; we schedule the work, hold the memory and rules, route jobs between agents, and keep everything reviewable. Your AI spend stays on your existing bill at your existing rate, and adding a fourth or tenth project does not multiply it.
What a Codex subscription costs
Codex is included in every ChatGPT tier, so if you already pay for ChatGPT you may already have what you need. These are OpenAI's prices, not ours:
| Plan | Price | What it suits |
|---|---|---|
| ChatGPT Go | $8per month | The cheapest way in. Light, occasional runs — a single project ticking over. |
| ChatGPT Plus | $20per month | The usual starting point. Several projects on a daily cadence. |
| ChatGPT Pro | $100–$200per month | Heavy, near-constant use. $100 buys the 5× tier, $200 the 20×. |
| ChatGPT Business | $20per user, per month | Shared workspaces where each teammate brings their own seat. |
OpenAI pricing, checked 29 July 2026 — see their current rates.
The floor and the ceiling
The floor is $8 a month. ChatGPT Go includes Codex, and that is genuinely enough to keep one project moving — a blog that publishes twice a week, an inbox triaged each morning. Add the cost of Honey Badgers itself and you are automating real work for less than a couple of lunches.
The ceiling is $200 a month. That is ChatGPT Pro at the 20× tier, the point where agents run more or less continuously across a lot of projects. There is nothing above it to buy on a subscription — past that you move to API billing and pay for use.
Most people land in the middle, on Plus at $20. The honest advice is to start at the bottom and move up when you actually feel the limit, because the jump from $20 to $100 is five times the money and the engine should earn it.
What different projects need
Cost tracks how hard your agents run, not how many projects you have on paper. A project publishing one post a week is nearly free to keep alive; one watching an inbox all day and drafting replies is not. Rough shapes:
- A publishing cadence. Research, draft, edit, image, publish — a handful of runs a week. Light, and the cheapest useful thing you can automate.
- Inbox triage. Runs whenever mail lands, so cost follows your mail volume rather than a schedule you set. Moderate, and spiky.
- A standing research or monitoring job. Something checking a source on a tight loop. This is the one that quietly eats a tier, because it never stops.
We are not going to quote you a projects-per-tier number. It would be a guess dressed up as a fact, and you would hold us to it. Two projects with the same name can differ tenfold in how often their agents actually run. Start on the cheapest tier that sounds plausible, watch the job feed for a fortnight, and move up only if you hit the wall.
Mixing engines across projects
Engines are set per project and per agent, so nothing forces one choice on everything you run. The pattern that works: a cheap, fast engine on the grinding — summarising, sorting, first drafts — and a stronger one reserved for work that goes in front of a customer.
It also means you are never locked in. If a better or cheaper model appears next quarter, you repoint one project at it and leave the rest alone, rather than migrating a platform.
Using an API key instead
Subscriptions are the cheap path for steady, predictable work. API keys are better for bursty work, because you pay for what you consume rather than renting a ceiling. Connect a Claude or OpenAI key directly and Honey Badgers uses it the same way it uses a subscription.
Plenty of people run both: a subscription covering the everyday cadence, and a key for the occasional heavy job that would otherwise blow through the tier.
Where this gets you into trouble
Bringing your own AI means you choose the quality, and that cuts both ways. An $8 tier on a nuanced writing job will read like an $8 tier on a nuanced writing job. The platform cannot make a weaker model cleverer than it is.
This is why nothing publishes behind your back by default. Every plan, draft and result comes through the Plan and Publisher feeds for you to approve, redirect or reject — so a bad engine choice costs you a review cycle, not your reputation. Once a project has produced good work for a while, you can let it run further on its own.
Not sure which tier your work needs?
Tell us what you want automated and we will tell you honestly what it would take to run — including whether it is worth automating at all.