App hosting is the part of OpenAI’s new developer plans that nobody has confirmed yet. On 23 September 2026 TestingCatalog published a screenshot of a new onboarding screen on OpenAI Platform headed “Choose your plan”, offering three tiers called Free, Prototype and Accelerate. The screen is specific: Prototype starts at $5, Accelerate starts at $50, and a credit box explains what $50 buys in tokens. What it does not say anywhere is that OpenAI will run your application for you.
That gap matters because the headline most people saw, plans “to enable developers to build and host their AI apps”, reads app hosting into a screen that only sells API access, Codex and Playgrounds. TestingCatalog’s own write-up is careful on the point: the discovered strings “confirm the new plan structure, not an OpenAI-hosted app deployment product”. DevDay is on 29 September in San Francisco, so the answer is six days away. Our AI models and tools hub tracks the releases around it.
This article reads the plan screen line by line, maps each plan onto the usage tiers OpenAI already runs, works through the token arithmetic on the $50 credit, sets out what OpenAI already hosts for developers, and lists the questions DevDay has to answer before anyone plans an app hosting move around it.
Table of contents
- What the OpenAI Plan Screen Actually Shows
- Free, Prototype and Accelerate, Plan by Plan
- Is This Really App Hosting? What the Evidence Says
- How the Plans Map Onto OpenAI’s Existing Usage Tiers
- What $50 Buys: Reading the Token Estimate
- What OpenAI Already Hosts for Developers
- Why OpenAI Would Want to Own App Hosting
- What App Hosting on OpenAI Would Mean for Developers
- What to Watch at DevDay on 29 September
- How to Prepare Before OpenAI App Hosting Arrives
- Frequently Asked Questions About OpenAI App Hosting
- References
What the OpenAI Plan Screen Actually Shows
The evidence is one screen and a seven-second clip, so it is worth reading exactly rather than summarising.
Where the screen appeared
TestingCatalog’s Alexey Shabanov published the find at 09:16 UTC on 23 September. The screen carries the OpenAI Platform header, a dollar-sign icon and the heading “Choose your plan”, with the line “Start experimenting for free or unlock more usage to build your idea.” A progress indicator at the top suggests it is the first of two onboarding steps, so there may be more to the flow than anyone has seen.
What the three cards say
Each card carries a one-line purpose. Free is to “Experiment with the API to explore your idea”. Prototype is to “Build lightweight prototypes to evaluate the API”. Accelerate is to “Launch your apps to production with our API”. All three sentences end on the API. None of them mentions deploying, serving or hosting an application, which is the first reason to read the app hosting claim cautiously.
The credit box underneath
Below the cards sits an “API credits” field pre-filled with $50 and a note: “$50 can cover about 250 million input or 40 million output tokens. Ideal for larger workloads and production apps.” That is a prepaid credit, the same mechanism OpenAI’s API billing already uses, not a subscription for compute or app hosting capacity.
No date and no announcement
No rollout date is visible in the flow, and OpenAI has not announced the plans. Runtime Wire, which followed up the report, lists the open questions: whether $50 is monthly or one-off, what usage each tier includes, and whether any plan bundles app hosting, compute or other services beyond API access.
Free, Prototype and Accelerate, Plan by Plan
Each card lists four short lines. Taken together they describe a ladder from a free trial to production traffic.
Free
The Free card lists “Free”, “Limited access to certain models”, “Usage limits per minute / day”, and “Start building for free with the API in Codex and limited access to Playgrounds”. The Codex line is the interesting one. It implies a developer can start building inside OpenAI’s coding agent without adding a card, which lowers the first step considerably.
Prototype
Prototype is “Starting at $5” with “Access all of our latest models”, “Enough usage to power prototypes and small apps”, and fuller Codex and Playgrounds access. “Small apps” is the closest the screen comes to describing something that runs in public, and even that line describes usage, not app hosting.
Accelerate
Accelerate is “Starting at $50” with all the latest models, “Higher usage & rate limits to power larger, production apps” and the same Codex and Playgrounds line. It was the card pre-selected in the screenshot, which is the usual pattern for the plan a vendor most wants chosen.
Playgrounds, plural
Every card says Playgrounds, not Playground. OpenAI has had a single Playground for prompts for years, so the plural hints at several workbench-style surfaces. That is a reasonable guess rather than a fact, and it is one of the things DevDay should settle.
| Plan | Card promise | Price shown | Models | Usage line |
|---|---|---|---|---|
| Free | Experiment with the API | Free | Certain models, limited | Limits per minute and per day |
| Prototype | Build lightweight prototypes | Starting at $5 | All latest models | Enough for prototypes and small apps |
| Accelerate | Launch apps to production | Starting at $50 | All latest models | Higher usage and rate limits |
Is This Really App Hosting? What the Evidence Says
The question that decides how much this matters is whether the plans include app hosting at all. The evidence points both ways.
The case for reading it as app hosting
TestingCatalog’s standfirst says the plans exist “to enable developers to build and host their AI apps”, and the article argues the wording “suggests OpenAI wants Platform to become a broader development environment rather than primarily an API dashboard”. The phrases “small apps” and “production apps” appear on two of the cards. Read generously, a ladder that ends in “Launch your apps to production” sounds like the front door to app hosting.
The case against
Read literally, every card ends with “the API”. The prices are “starting at” figures that match API credit top-ups. The same TestingCatalog article concedes that “It remains unclear whether developers will eventually be able to deploy or host complete applications directly there.” Runtime Wire makes the same point more bluntly: it is not known whether general app hosting will be available at all.
The honest summary
The plans are confirmed by a screenshot. App hosting is an inference from the direction of travel. Treat any budget, architecture or migration decision that assumes OpenAI app hosting as premature until OpenAI itself describes a deployment product, its limits and its price.
Why the distinction matters in practice
A prepaid API plan changes how you pay for model calls. An app hosting product changes where your code, your users’ data and your uptime live. The second is a much larger commitment, with contracts, data residency and exit costs attached, and it deserves its own evaluation rather than a decision made on the strength of a plan name.
| Question | Confirmed by the screen? | What is known |
|---|---|---|
| Three plans named Free, Prototype, Accelerate | Yes | Shown on the onboarding screen |
| Prices | Partly | “Starting at” $5 and $50; billing period not stated |
| Model access | Yes | Limited on Free, all latest models on paid plans |
| Codex and Playgrounds | Yes | On all three plans, limited on Free |
| App hosting or deployment | No | Not mentioned anywhere on the screen |
| Rollout date | No | None visible; DevDay is 29 September |
How the Plans Map Onto OpenAI's Existing Usage Tiers
The prices on the cards look new. They are not.
The tiers OpenAI already runs
OpenAI’s rate-limit documentation already sorts every API organisation into usage tiers that rise automatically as spend goes up. The free tier requires an allowed geography and carries a $100 monthly usage limit. Tier 1 needs $5 paid, Tier 2 $50, Tier 3 $100, Tier 4 $250 and Tier 5 $1,000, with monthly usage limits of $100, $500, $1,000, $5,000 and $200,000 respectively.
The prices are the existing thresholds
Put those next to the plan screen and the “starting at” figures stop looking new. Prototype’s $5 is exactly the Tier 1 qualification. Accelerate’s $50 is exactly the Tier 2 qualification. The simplest reading is that OpenAI is putting friendly names and a guided onboarding on thresholds that already exist, rather than publishing an app hosting price list.
What the renaming changes
Names change behaviour. A developer who sees “Tier 2” sees a rate limit; a developer who sees “Accelerate: launch your apps to production” sees a product. That framing, plus Codex on the Free plan, is how an API dashboard starts to feel like a development platform before any app hosting exists.
What it does not change
If the mapping holds, the money still buys tokens, and the limits still rise with cumulative spend. Nothing on the screen suggests a flat fee that includes compute, bandwidth or storage, which is what app hosting normally costs.
| Usage tier today | Qualification | Monthly usage limit | Matching new plan |
|---|---|---|---|
| Free | Allowed geography | $100 | Free |
| Tier 1 | $5 paid | $100 | Prototype, “Starting at $5” |
| Tier 2 | $50 paid | $500 | Accelerate, “Starting at $50” |
| Tier 3 | $100 paid | $1,000 | No card; above Accelerate’s floor |
| Tier 4 | $250 paid | $5,000 | No card; above Accelerate’s floor |
| Tier 5 | $1,000 paid | $200,000 | No card shown |
The step from Free to Tier 4 multiplies the monthly ceiling fifty times, and the jump to Tier 5 is another forty times on top.
What $50 Buys: Reading the Token Estimate
The credit note is the only price-per-token information on the screen, and it does not match the price list.
The implied rates
Divide the credit by the token counts on the screen and you get the rates OpenAI used for the estimate. $50 over 250 million input tokens is $0.20 per million. $50 over 40 million output tokens is $1.25 per million.
They match no single GPT-6 price
Neither figure matches a current GPT-6 list price. GPT-6 Luna costs $0.10 per million input and $0.50 per million output, so $50 would buy 500 million input or 100 million output tokens. GPT-6 Sol costs $2 and $10, so $50 buys 25 million input or 5 million output. The estimate sits between them: twice Luna’s input rate and a tenth of Sol’s.
The likely explanation
The note may be a blended estimate across models, or it may have been written against an older price list and not updated after the 22 September cuts. Either way it is marketing copy on an unreleased screen. Budget from the published price list, not from the onboarding blurb.
What that means for a small app
At Luna’s list price, a chatbot turn using 2,000 input and 500 output tokens costs $0.00045, so $50 covers roughly 111,000 turns. At Sol’s price the same turn costs $0.009, about 5,556 turns for $50. Model choice moves the bill twentyfold. App hosting, if it arrives, would be a separate line on top of either figure.
Input tokens bought by $50 fall by a factor of twenty between the cheapest and the mid-tier model.
What OpenAI Already Hosts for Developers
OpenAI already runs compute on developers’ behalf. It just does not call it app hosting.
The Agents API
On 10 September OpenAI opened the Agents API as a public beta. It runs cloud agents on an OpenAI-managed Codex harness inside OpenAI-hosted sandboxes, as we covered in our GPT-Live-1 and Agents API write-up. That is hosted execution, scoped to agent tasks.
Hosted containers
The API price list already bills Hosted Shell and Code Interpreter container sessions by the minute, with a five-minute minimum per session. That is metered compute, the raw material of app hosting, sold today inside the API rather than as a product of its own.
Codex on every card
Codex is OpenAI’s coding agent. Putting it on every plan card, including Free, suggests the onboarding is designed to take someone from an idea to working code inside OpenAI’s own tools, which is the step before app hosting.
What is still missing
None of this serves a public web application with its own domain, database and users. Agents and containers run tasks; app hosting runs a product continuously. That last step is the one TestingCatalog could not find evidence of, and our earlier report on the Managed Agents leak before DevDay did not find it either.
Why OpenAI Would Want to Own App Hosting
The commercial logic for app hosting is strong, which is partly why the headline ran ahead of the evidence.
Control of the runtime
TestingCatalog notes that the Agents API “already gives OpenAI more control over the runtime layer behind developer applications”. Owning app hosting would complete that stack: model, tools, code and runtime from one vendor.
More revenue per developer
An API customer pays for tokens. An app hosting customer pays for tokens plus compute, storage and bandwidth, month after month, and is much less likely to switch model provider when a rival cuts prices.
The price war
OpenAI and Anthropic cut prices hard on 22 September, as our AI price war analysis covered. When token prices fall, owning more of the stack is the obvious way to protect revenue, and app hosting is the largest piece of the stack OpenAI does not yet sell.
A developer conference to fill
Sam Altman posted on 22 September: “We ❤️ developers. See you next week.” A guided path from Free to production, and possibly app hosting at the end of it, is exactly the kind of launch a developer conference is built around.
What App Hosting on OpenAI Would Mean for Developers
If OpenAI does announce app hosting, the benefits and the costs are both easy to predict.
Faster from prototype to production
For a solo developer or a small team, one account covering model access, a coding agent and app hosting would remove several integration steps and at least one bill. That convenience is real, particularly for internal tools and early prototypes.
Harder to leave
The trade-off is lock-in. Code written in a vendor’s environment, running on its app hosting and calling its models, is costly to move if prices or policies change. Keep model calls behind your own interface so that the provider can be swapped without a rewrite.
Data and residency questions
Our Agents API analysis found that the Agents API did not qualify for Zero Data Retention or European processing at launch. Any OpenAI app hosting product would need clear answers on where user data lives, which matters for UK and EU businesses working under UK GDPR.
Security responsibilities
Hosted code still needs patching, secrets management, access control and monitoring. Ask which of those the vendor handles and which stay with you, and treat cybersecurity controls as a precondition of any app hosting move rather than a later task.
What to Watch at DevDay on 29 September
DevDay is the obvious venue, and it will answer most of the open questions in one afternoon.
A deployment product, named and priced
The single most important signal is whether OpenAI names an app hosting or deployment product at all, and whether it has its own price list separate from tokens.
The billing period for $5 and $50
“Starting at” could mean a one-off credit purchase, as today, or a monthly minimum. The difference matters for budgeting, and the screen does not say.
What happens to the numbered tiers
If Free, Prototype and Accelerate replace Tier 1 and Tier 2 in the dashboard, check whether the monthly usage limits change with the names.
Data controls and regions
For any app hosting announcement, look for data retention terms, European processing and the list of regions before anything else.
How to Prepare Before OpenAI App Hosting Arrives
Nothing on the screen requires action today, but a few habits make any announcement easier to evaluate.
Budget from the published price list
Use the per-million rates on OpenAI’s pricing page for estimates, not the onboarding note. The two already disagree.
Abstract your model calls
Route every model call through one internal function or service. It costs little now and makes any future choice about app hosting or model provider reversible.
Decide hosting on its own merits
Where an AI product runs is a separate decision from which model it calls. If you are weighing that choice for a new product, our software development team can help you compare the options.
Watch your usage tier
Check which tier your organisation sits in today. If the plan names arrive, you will know immediately whether your limits moved.
Frequently Asked Questions About OpenAI App Hosting
Has OpenAI announced app hosting?
No. The evidence is an onboarding screen found by TestingCatalog that names three plans. None of the plan cards mentions hosting or deploying an application.
How much do the new plans cost?
The screen shows Free, Prototype “Starting at $5” and Accelerate “Starting at $50”. The billing period is not stated.
Are the plans live?
Not publicly. No rollout date is visible, and OpenAI DevDay on 29 September is the likely venue for any announcement.
Do the plans replace the usage tiers?
Unknown. The $5 and $50 prices match the current Tier 1 and Tier 2 qualifications exactly, which suggests a renaming of existing thresholds.
What does $50 buy?
The screen says about 250 million input or 40 million output tokens. At GPT-6 Luna’s list price it would buy 500 million input or 100 million output tokens.
References
DevDay: New Plans and new platform for building AI apps
OpenAI’s $50 developer plan appears in an onboarding flow ahead of DevDay
Rate limits and usage tiers, OpenAI API documentation
API pricing, OpenAI API documentation
OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes
Anthropic and OpenAI roll out cheaper models in first release since call for slowdown
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
