For $20 a month, your ChatGPT Plus subscription now buys what looks like $100–$200 worth of tokens, and OpenAI is (for now) fine with you spending them inside third-party harnesses instead of its own apps. Here’s why that’s a great deal, and why it might not last. A few weeks ago, Tibo Sottiaux, an executive at OpenAI, announced that you can use your ChatGPT account (the paid OpenAI subscription that starts at $20/month with Plus) inside third-party harnesses. He added that Pi and OpenCode, two popular open-source coding agents, already make up 10% of Codex traffic.
The move was well received, especially with Anthropic going the other way after banning non-official harnesses from its subscription. And it’s a genuinely good deal: a $20/month plan is worth somewhere around $100–$200 in tokens, by OpenAI’s own estimate. (If you want to try it, here’s how to run Claude Code on a ChatGPT Plus subscription.)
Not bad, right? But is it going to last?
Not much, and that’s exactly what makes it fragile. Nothing in OpenAI’s terms of use explicitly permits or prohibits using your ChatGPT subscription inside a non-OpenAI tool. It reads like something they tolerate, not something they’ve committed to.
What they do say plainly is that they reserve the right to terminate your access if your use could cause risk or harm to OpenAI, its users, or anyone else.
And economic harm is an easy argument to reach for. Subsidizing tokens for people who never touch OpenAI’s own tools puts pressure on a business that isn’t profitable yet, and that pressure only grows if this gray-zone usage keeps climbing.
The underlying economics are the same as Anthropic’s subscription. At the end of the day, both companies have to turn a profit.
The key difference with Anthropic is that OpenAI has publicly embraced third-party harnesses on a ChatGPT subscription, more than once, from Sam Altman on X to Tibo’s recent post.
They also shipped a “Sign in with ChatGPT” flow that covers subscription users, not just API customers. And their Codex for Open Source program is even clearer:
Developers should code in the tools they prefer, whether that’s Codex,
[OpenCode],[Cline],[pi],[OpenClaw], or something else, and this program supports that work. None of these statements is contractual. But the direction is unmistakable. Maybe they lean into it to compete with Anthropic, maybe it’s genuine conviction. Either way, the signal is strong.
We’re seeing a lot of well-crafted third-party harnesses ship, most accepting bring-your-own-key (BYOK) and, increasingly, OAuth subscription sign-in too.
The coding space is especially crowded: OpenCode, Droid, OpenHands, KiloCode, Crush, Aider, and Pi, to name a few. Autonomous agents like OpenClaw and Hermes are hugely popular and burn through tokens. Agent orchestrators like Paperclip and Conductor are a legitimate category now too, and we’ll likely see more vertical harnesses soon.
Why the long list? Because token consumption from these tools is growing far faster than official ChatGPT and Codex usage. What happens when third-party harnesses go from 10% of Codex traffic to, say, 50%? Does OpenAI keep footing the bill?
OpenAI needs a strong reason to keep subsidizing that many tokens: funneling you toward its own tools, or staying on the good side of the communities it wants to win. But in the end it comes down to the numbers, and with an IPO on the horizon, who knows how long this lasts?
Our advice? Enjoy it while it lasts.