Your users can pay your AI bill with ChatGPT. OpenAI keeps the meter.

The short version

• On September 29, spending your ChatGPT plan in other apps became an official program. At DevDay, OpenAI let Plus and Pro subscribers route AI requests in 15 partner apps to their plan, with Lovable to follow, and each user can set a weekly limit per app. Coding tools make up most of the list, from Devin and Warp to the open-source OpenCode.
• The pitch lands hardest on small AI products. Sam Altman told developers the change means “now you don’t have to cover their token costs to get them going.” If you read our study of why flat-priced AI wrappers lose money, that is the heavy user’s bill moving to the heavy user.
• Open source gets it now, and paid apps get a waitlist. OpenAI’s developer docs open plan usage to all open-source partners and to private clients OpenAI selects. A paid or remotely hosted app has to apply and wait for a place in a limited trial.
• It is one allowance, and everything draws on it. Connected apps spend the same ChatGPT Work and Codex usage the subscriber spends directly. On Plus, the five-hour limit covers all of them, and “no app receives a separate allowance.” OpenAI estimates that window at 15 to 160 local messages on GPT-6.1 Sol, with cloud tasks using more.
• The same day, OpenAI changed what a plan buys. Pro 200 now includes less usage for new sign-ups, and current subscribers switch after October 29. The Next Web reports Codex and Work usage falling from 20 to 10 times the Plus allowance, next to a new $500 tier.
• Agent storefronts opened the same month, built for discovery more than payment. Meta, Anthropic and OpenAI all widened their directories for outside developers. Anthropic’s and OpenAI’s new purchase paths serve enterprise budgets, OpenAI announced no app-store-style billing or revenue share, and Meta points developers to Stripe.
• The argument: letting users pay with their ChatGPT plan can take inference off a solo builder’s books, but it prices your product’s capacity in an allowance OpenAI sets and can shrink. Offer it as one way to pay, with your own checkout and API key still behind it.

What shipped on September 29

Sign in with ChatGPT is older than DevDay. OpenAI released it as a beta in July with partners including Airtable, GitLab, HubSpot, Notion, Supabase and Vercel, and at that point it was mostly a login button. What changed at DevDay is the second permission. When a Plus or Pro user signs in to a participating app, they can now choose to let that app’s AI requests count against the usage included in their ChatGPT plan.

The user sees a consent screen, then a line in ChatGPT’s settings for each connected app, where they can cap it at a percentage of their weekly usage. No API key changes hands. Signing in shares the user’s name, email address and profile picture, and using the plan does not by itself give the app their conversations or memories. If the plan runs out, the app can draw on the user’s ChatGPT credits only if the user has turned that on. And OpenAI’s help page is careful on one point: “The app may charge separately for its subscription, infrastructure, services, or premium features.”

The launch list says who this is for. OpenAI’s directory lists twelve apps for plan usage: eleven live at launch (Amp Code, Conductor, Dactyl, Devin, Hermes Agent, Hyperagent, Kilo Code, Notion, Vercel, Vorflux and Warp) and Lovable, marked coming soon. Most of them are coding tools. Four open-source clients round out the sixteen: OpenCode, OpenClaw, Pi and T3.

Some of this was already happening. Dactyl, a vibe-coding platform for native mobile apps, was showing the option in late August, and OpenClaw announced in May that users could power agent turns with a ChatGPT subscription through Codex login. DevDay folded those routes into one program the user manages from ChatGPT, and OpenAI told The New Stack it will add partners quickly.

Kilo’s launch post shows what it looks like from the inside. The plan works across its editor extensions, command line, cloud agents, code reviews and mobile app, and in Kilo’s words “the ‘key’ is the ChatGPT subscription you already pay for.”

Why this matters to a small AI product

The hardest line on a solo builder’s AI product is the one nobody controls: what the heaviest user costs. Price at $20 a month and the person running your agent around the clock can cost you more than they pay. Our wrapper study found the big AI products answering that with meters and caps.

Plan usage answers it differently. The heavy user’s requests come out of the heavy user’s own ChatGPT allowance, and when that runs out, OpenAI’s limit is the one they hit. Free trials get cheaper to offer, because a new user who brings a plan costs you almost nothing in inference. That is the part of Altman’s pitch that holds up.

It also makes your product’s capacity someone else’s number, and agentic work eats an allowance fast. When we ran OpenAI’s Codex on the Free plan for our Proof, one app build used 37 percent of that plan’s 30-day meter. Free has its own meter and cannot be shared with other apps, but the build shows the appetite. A Plus subscriber who connects your tool and two other coding agents, and also codes in Codex, is splitting a single five-hour window four ways.

What OpenAI still controls

Five things stay on OpenAI’s side of the table, and each one reaches into your product.

• Who gets in. Open-source clients can register themselves today from the developer docs, and closed-source clients are chosen by OpenAI. A paid or hosted app can fill in an interest form, which puts it on a waitlist, and the terms for commercial token sharing are not published.
• The meter. The user sets your app’s weekly share, from 10 to 100 percent in OpenAI’s example. On Plus, every connected app shares one five-hour window with the subscriber’s own Codex and ChatGPT Work sessions. Pro plans currently have no five-hour limit.
• The stop. When the user hits a limit, their plan’s or the weekly cap they set for your app, and has not allowed credits, the request fails with a usage-limit error. OpenAI does not silently switch it to another billing path, so unless you offer a paid option of your own, your product stops for that user until their limit resets.
• The feature list. During the preview, plan mode does not support image generation, file search, Code Interpreter, native computer use or hosted connectors. A feature built on any of them needs another billing path behind it, such as your own API key or the user’s.
• The front door. OpenAI’s docs ask for a Continue with ChatGPT button, placed with your other sign-in options at comparable visual prominence and in approved OpenAI branding. A website may also use an approved Sign in with ChatGPT format. The customer’s identity arrives from OpenAI.

The $200 plan got lighter the same day

While OpenAI was inviting developers to build on its plans, it changed what one of them buys. New Pro 200 subscriptions come with a lower allowance, set, OpenAI says, “to reflect our increasingly efficient models.” Existing subscribers keep the old allowance through October 29, and the price stays $200. The Next Web reports that from October 30 Pro 200’s Work and Codex usage falls from 20 to 10 times the Plus allowance. Tibo, who works on Codex and ChatGPT at OpenAI, previewed the change in dollars the night before: the new plan “will net out at half the dollar in API spend compared to the old Pro $200 plan.” He argued subscribers will still get more work done than on Pro 200 a month earlier, pointing to OpenAI halving the price of GPT-6 Sol and GPT-6 Luna that week. Existing subscribers also get a one-time $2,500 usage-credit grant that expires at year end, The Next Web reports. For the heaviest users, OpenAI added Pro 500, at 25 times Plus.

The change was announced for Pro 200 alone. It still shows who sets the exchange rate. OpenAI decides both what a plan includes and what the API charges, and a product that tells its users their ChatGPT plan covers the AI inherits whatever OpenAI decides a plan is worth.

The storefronts are open. Bring your own checkout.

September also opened three storefronts for agents. Meta opened its Muse agent to outside connectors; its platform page promises each one a review “for functional, security, and legal requirements” plus end-to-end testing, and editors choose which to feature. Alexandr Wang, Meta’s chief AI officer, said on stage at Connect that more than 1,500 applications arrived in under a week.

Anthropic launched the Claude Marketplace on September 23 with more than 2,000 connectors and plugins. Two days later it opened a plugin submission portal to developers on paid Claude plans. Submissions get a safety scan, and live plugins get analytics on installs and the searches that find them. At DevDay, OpenAI said ChatGPT will start suggesting apps in the middle of a conversation when one fits the task, and that developers will be able to track a plug-in’s review and request a human reviewer.

Payment is thinner. The buying routes Anthropic and OpenAI built serve enterprise budgets: teams can put part of their committed Anthropic spend toward listed products, and OpenAI’s new marketplace lets eligible enterprise customers apply part of their OpenAI commitment to software from its first 32 partners. TechCrunch noted that OpenAI announced no billing or revenue-sharing system comparable to an app store’s. Meta has gone furthest, with a Stripe partnership for accepting payments through Link, and Mark Zuckerberg said Muse is free for a huge number of tokens “with the expectation that over time we will profit by taking a small fee from transactions.”

So the storefronts supply reach, and Sign in with ChatGPT supplies inference. Collecting your own price is still your job.

What to do this week

1. If you ship a free, open-source tool that runs on the user’s machine, add plan usage now. The developer docs cover registration, and removing the API-key step is a real gain for new users. Keep bring-your-own-key as a second path.
2. If you run a paid or hosted app, join the waitlist and keep your billing. Plan usage is one way users can choose to pay. Keep your own checkout working for everyone else.
3. Build for the stop. Catch the usage-limit error and tell the user they have hit a ChatGPT usage limit, either their plan’s or the cap they set for your app. Make a Manage usage link to their ChatGPT settings the main action, and offer your own paid route second.
4. Say what you cannot promise. If your AI runs on the user’s plan, your speed and volume depend on their plan and their other apps. Put that in your onboarding in plain words.
5. If you are the buyer, set caps. Connect your plan to the coding tools that support it, cap each one’s weekly share, and check which of them still charge their own subscription on top.