Marketing

AI Wrapper: What It Is, What It Costs and What New AI Tools Name

/18 min read

An AI wrapper is a product built on an AI model someone else trained, reached through that maker's API, with its own interface, prompts, data or workflow around it. On 19 November 2026 we read the home pages of the newest tools on three AI tool directories: of 347 that answered, 180 (52%) named a third-party model or model maker, 128 named two or more and two claimed a model of their own.

Below: the AI wrapper meaning and how a wrapper differs from a model maker and an agent, those 347 tools as AI wrapper examples, what a GPT wrapper, an LLM wrapper and a Claude wrapper are, whether ChatGPT and Perplexity count, what each request costs on 12 models priced from their makers' own pages, whether a wrapper is a business and where new AI products get listed.

Short version

An AI wrapper is a product on someone else's model. A little over half of new AI tools name such a model or its maker on their home page, 180 of 347, OpenAI most often, though in a sample of those pages the name was more often a plug-in or a passing mention than the engine. The same work can cost 80 times more on one model than on another, and the model is rented, so a wrapper's moat is its data, its workflow and its reach. For reach, AI tool directories are where new AI products get listed: AI Tools Up charges $19 to skip its queue, and $29 at Submitator covers 40 directories you pick from the 202 in our catalog.

What is an AI wrapper? The meaning

What is meant by a wrapper? In software, a wrapper is code that sits around something else and gives it a simpler or more specific face. So what is a wrapper in AI? It is a product whose core is a call to a model the company did not train: it sends the user's input, plus its own instructions, to a model such as GPT, Claude or Gemini, and shows the answer in its own interface. The AI wrapper meaning most people intend is exactly that: an app, often a wrapper app on the web or a phone, that does one job with a general model underneath.

The word is often used as a put-down, as in "it's just a wrapper". It describes how a product is built, not whether it is any good. A tax tool that calls a model and checks the answer against your own records is a wrapper; so is a chat box with a logo. The difference between them is everything around the call. At its smallest, the call looks like this:

POST https://api.example.com/v1/messages
{
  "model": "the-model-you-picked",
  "system": "You write short, polite follow-up emails.",
  "input": "Customer asked about pricing on Tuesday, no reply since."
}

The address and field names above are placeholders; each maker's API names them its own way. The product is the system prompt, the input it collects, what it does with the reply and the screen around it.

AI wrapper vs model maker vs agent

A model maker trains the model and sells access to it. A wrapper buys that access per request. An AI agent wrapper adds a loop: the model chooses a tool, reads what came back and decides again, so one task can be many requests. Harness is the word some builders use for that loop and the code that keeps it running (tools, memory, retries); DeepSeek's API docs, for one, list a DeepSeek Harness among their agent integrations. On AI harness vs wrapper, then: the harness is the loop around a model, and when that model is someone else's, the product as a whole is a wrapper that runs it in steps rather than once.

What a model maker, an AI wrapper and an AI agent each do

What a model maker, an AI wrapper and an AI agent each do
 Model makerAI wrapperAI agent
Who trained the modelThe company itselfSomeone else, called through an APIUsually someone else
What it addsThe model, and often its own appAn interface, prompts, data or a workflow for one jobA loop: the model picks a tool, reads the result, decides again
What it pays per requestIts own computeThe maker's price per million tokensThe same, once per step of the loop
What it can switchNothing below itThe model, by changing one API callThe model, and the tools it is given

Our definitions, not a measurement. A product can sit in two columns: an agent built on someone else's model is a wrapper with a loop, and a model maker's own app is not a wrapper of anything.

AI wrapper examples: 347 new AI tools, read for the models they name

Lists of AI wrapper startups name the famous ones. We counted the new ones instead. On 19 November 2026 we took the newest tools on three AI tool directories we submit to: DangAI's list of new AI tools (its ten newest pages, reaching back to 8 November 2026), every product under Turbo0's AI tag published from 1 to 19 November and the latest products on The AI Tools Dir. That gave 358 tools after 12 repeats, and 347 home pages answered. For each we read the page the server sends and asked four questions. Does it name a model or model maker? Does it name more than one? Does it say it is powered by or built on a model? Does it claim a model of its own?

180 of the 347 (52%) named a third-party model or maker somewhere on the page. 128 named two or more, so more than two thirds of the tools that name one name several; 40 named five or more, up to 12. A phrase like powered by, built on or uses sat next to a model's name on 17 pages, about one in twenty. Two pages claimed a proprietary model: one a video maker that names outside makers too, the other a trading tool whose proprietary models are trading models rather than AI. The remaining 167 named no maker at all.

AI wrapper examples, counted: 347 new AI tools on three AI directories, 19 Nov 2026

AI wrapper examples, counted: 347 new AI tools on three AI directories, 19 Nov 2026
DirectoryIts newest listToolsSites readName a model makerName two or morePowered-by phraseOwn model claim
DangAIDR 83New AI tools, ten newest pages (listed 8 to 19 Nov)237229108 of 22978121
Turbo0DR 81AI tag, published 1 to 19 Nov10910868 of 1084751
The AI Tools DirDR 58Latest Products block12104 of 10300
All three 358347180 of 347 (52%)128 of 347 (37%)172

Every tool on three AI tool directories we submit to, read on 19 November 2026: DangAI, its New AI tools list at /new, pages 1 to 10 (24 tools a page, newest first), each tool dated by the Listed line of its own page and linked to its website by the tool record on that page; Turbo0, every product under Turbo0's own AI tag published from 1 to 19 November 2026, newest first (its category list filtered by that tag and those dates, pages 1 to 11), each dated and linked by its item page's Published date and Visit Website button; The AI Tools Dir, the Latest Products block of its home page, every product in it, linked by each card's Visit button. Why these three: before any tool's home page was read, 58 AI tool directories in our catalog were checked for a list of their newest AI tools, in order, in the HTML they serve, with each tool linked straight to its own site. DangAI's list and The AI Tools Dir's block met that; Turbo0 lists products of every kind, so its own AI tag, filtered by publish date, was taken in place of two AI tool lists that reach each tool through a click-tracking redirect or a script. The others were left out because their lists mix products of every kind, are not newest first, need a click to load more or reach the tool through a click-tracking redirect or a script. Lists and listing pages read between 21:04 and 21:09 UTC, the tools' home pages between 21:10 and 21:12 UTC. Frame: every tool on the three lists as read, in the order DangAI, Turbo0, The AI Tools Dir; a tool on two lists (the same host, www stripped) counts once, on the first (12 Turbo0 listings were tools already on DangAI's list). Read: one GET of the tool's website as the directory links it (tracking parameters removed), redirects followed, no JavaScript; answered when it returned the tool's own page; a bot wall (a 403 challenge titled Just a moment or Attention Required) leaves the base; a DNS or TLS failure, a timeout, a suspended account or another error is not answered; 8 sat behind a bot wall and 3 did not answer (a suspended account, a 502 and a connection that failed). Seven home pages served an empty shell: a page whose visible text was under 200 characters is a shell and was read again in a headless browser that loads it, waits 3.5 seconds and clicks nothing. Text read: the page's title, meta description, og:description and visible text; never scripts, JSON payloads or links. Names a maker: the visible copy names a model or model maker from our list (OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, xAI, Alibaba, Stability AI, Black Forest Labs, Midjourney, ElevenLabs, Kuaishou, Runway, Suno, Udio, ByteDance, MiniMax, Luma AI, Moonshot AI and Cohere, each by the model names its own pages use). Two or more: the page names two or more of those makers. Powered-by phrase: one of the phrases powered by, built on, built with, runs on, running on, based on, uses, using, leverages, driven by, backed by, fueled by, integrates, integrated with or access to stands within four words before a model's name; read in context, five of the 17 matches were not about the tool's own models (a prompt to use it from ChatGPT, a showcase item's name, the reader's own OpenAI SDK, a model list the phrase ran into and a question about how it differs from ChatGPT). Own model: the page says our own model, proprietary model, proprietary AI, in-house model, trained in-house or the like; one of the two matches was about trading calculations, not AI. The rules were written before any home page was fetched, and the counts are by rule. No tool from a directory is named on this page.

Which models AI wrapper companies name

OpenAI leads by a distance: 146 pages named it, through ChatGPT, GPT or OpenAI itself, and of the 52 pages that named one maker alone, 31 named OpenAI. Anthropic (93), Google (76) and xAI (Grok, 57) follow. Then the video model makers: ByteDance (Seedance and Seedream, 51), MiniMax (Hailuo, 32) and Kuaishou (Kling, 30) each turned up more often than DeepSeek (16), since a video tool tends to list every model it offers.

Which model makers new AI tools name: 347 home pages, 19 Nov 2026

Which model makers new AI tools name: 347 home pages, 19 Nov 2026
Model makerPages naming itShare of 347 read
OpenAI14642%
Anthropic9327%
Google7622%
xAI5716%
ByteDance5115%
MiniMax329%
Kuaishou309%
Alibaba206%
DeepSeek165%
Black Forest Labs113%
Suno93%
ElevenLabs62%
Meta51%
Mistral51%
Moonshot AI51%
Udio51%
Midjourney31%
Luma AI21%
Runway21%

A page counts once for each maker it names, so the column adds up to more than 347. The names that counted for each maker: OpenAI (OpenAI, ChatGPT, GPT and its versions, GPTs, DALL-E, Sora, Whisper), Anthropic (Anthropic, Claude), Google (Gemini, Gemma, Veo, Imagen, Nano Banana), xAI (Grok, xAI), ByteDance (Seedance, Seedream), MiniMax (MiniMax, Hailuo), Kuaishou (Kling), Alibaba (Qwen, Wan 2.x), DeepSeek (DeepSeek), Black Forest Labs (FLUX, Flux.1, Flux Kontext, Flux Pro, Flux Dev, Flux Schnell, Black Forest Labs), Suno (Suno), ElevenLabs (ElevenLabs), Meta (Llama, Meta AI), Mistral (Mistral, Mixtral, Codestral), Moonshot AI (Kimi), Udio (Udio), Midjourney (Midjourney), Luma AI (Luma AI, Dream Machine, Luma Ray), Runway (Runway ML, Runway Gen, Runway Aleph). Also looked for and not found on any page: Stability AI and Cohere. Kuaishou makes Kling, ByteDance makes Seedance and Seedream, MiniMax makes Hailuo. A page that names a maker may run its model, connect to its app or mention it in passing; the reading of 30 below splits them.

AI wrapper apps by what they make

Of the kinds the rule names, video is the biggest: 93 of the 347 tools make video, and 55 of them name a model maker, most of them several, since a video app often lets you pick the model per clip. Agents name a maker most often, 27 of 36 (75%). Image tools name one least often of the named kinds, at 39%.

AI wrapper apps by what they make: how often each kind names a model

AI wrapper apps by what they make: how often each kind names a model
What the tool makesToolsName a model makerName two or more
Video9355 of 93 (59%)43
Image4618 of 46 (39%)12
Agent3627 of 36 (75%)22
Text3518 of 35 (51%)11
Code, apps and APIs2012 of 20 (60%)9
Audio and voice158 of 15 (53%)5
Other (none of the above in its summary)10242 of 102 (41%)26

Kind of output: from the title, meta description and first heading alone, the first that fits of agent, code (including app and website builders and APIs), video, image, audio and text; other when none fits. The words behind each kind are in the census rules; a tool whose summary names none of them is "other". Counts are of the 347 home pages read on 19 November 2026.

A model's name is not proof of a wrapper

Naming a model and running on it are different things. We drew 30 of the 180 naming pages at random and read each mention where it sits. Eleven named the model as the tool's engine: a model menu, a line saying which model makes each clip. Seven named a model's app they plug into, such as an MCP server for Claude or a ChatGPT plugin. The other twelve mentioned one in passing: an Ask ChatGPT button, a testimonial, a comparison, a customer's logo. So a home page that says ChatGPT tells you little, and a page that names no model may still run on one: the census counts what the pages say, not what the tools run.

What the name means: 30 naming pages read in context

What the name means: 30 naming pages read in context
The model's name isWhat it looked likeOf 30
The tool's engineA model menu, a line saying what each video or image is made with, a router that sends tasks to named models11
Something it connects toAn MCP server for Claude, a ChatGPT plugin, a prompt to run it from ChatGPT, a field for a Suno link7
A passing mentionAn Ask ChatGPT button, a quote from a ChatGPT forum, a comparison, a customer's logo, a footer badge, an AI engine the tool tracks12

30 of the 180 pages that named a maker, drawn at random (seed 20261119), each mention read where it sits on the page and put in one of three classes defined before the tally. A sample, so the split is a guide to the 180, not a count of them.

AI Tools Up: $19 to skip one queue. Submitator: $29 for 40 directories.

AI Tools Up is one of the 59 AI tool directories in our catalog, and so are the three above. Pick 39 more alongside it: a description written for each site, the badges kept live. 70 for $49, 110 for $79, paid once.

Start: from $29

What is a GPT wrapper? ChatGPT, LLM and Claude wrappers

A GPT wrapper is an AI wrapper built on OpenAI's GPT models: the product sends its requests to OpenAI's API and shows the answers in its own interface. In use, the GPT wrapper meaning is often dismissive: a ChatGPT wrapper (or chat GPT wrapper, as it is often typed) usually means a product a user could replace with a good prompt in ChatGPT itself. Whether that is true of a given product depends on what it adds; the term alone does not say.

LLM wrapper

An LLM wrapper is the same idea without naming the maker: a product on top of any large language model. Many name more than one: of the 180 naming tools in our census, 128 named two or more makers. A wrapper that can move from one model to another when prices or quality change owns more of its fate than one tied to a single API.

Claude wrapper

A Claude wrapper runs on Anthropic's Claude models. In the census, 93 of the 347 pages named Claude or Anthropic, second to OpenAI, and in our reading of a sample some named it because they offer an MCP server or a plugin that works from inside Claude, not because they call it.

API wrapper: the general term

What is an API wrapper? Outside AI it is a library that hides an API's raw requests behind simpler functions, so a developer calls one function instead of building the request by hand. Most AI wrappers use an API this way under the hood; the AI meaning is about the product, the general meaning about the code.

Voice AI wrapper

A voice AI wrapper does the same with speech models: it takes a call or a recording, sends it to a speech-to-text, text-to-speech or speech-to-speech model and handles the rest. Voice is often priced by the minute or by the character rather than by the token: xAI's models page, for one, prints “STS $0.08 / min” among its voice prices.

Is ChatGPT a wrapper? Is Perplexity?

Is ChatGPT a wrapper? No product needs the label to be judged, so here is what each company's own pages say it runs. ChatGPT is OpenAI's app on OpenAI's models. Its pricing page puts “Unlimited text chats with GPT-6” on the free plan and “Advanced reasoning models with GPT-6” on Plus, and its plan table names GPT-6.1 Sol, GPT-6 Astra, GPT-6 Sol and GPT-6 Luna: the same family OpenAI sells by the token in its API price list, where gpt-6-luna costs $0.10 per million input tokens. A product that calls those models and adds screens is a GPT wrapper; ChatGPT is the maker's own front end.

Perplexity is a different case. Its Pro page lists models from other makers, each with its own line: GPT-6.2 Sol (“OpenAI's versatile model”), Gemini 4 Neon (“Google's latest model”), Claude Sonnet 5.5 (“Anthropic's fast model”), Kimi K3 (“Moonshot AI's latest model. Hosted in the US.”), GLM 5.3 (“Z.ai's most advanced model. Hosted in the US.”) and Grok 4.7 (“xAI's most advanced model”), and it says “Perplexity orchestrates the best AI model for each task.” Its API docs say the same of the developer side: “The Agent API provides a unified interface to models from multiple providers with transparent token-based pricing.” The same page files one model under Perplexity itself, “Perplexity’s grounded search model”, and its embeddings page describes “The embedding models behind Perplexity Search”: models of its own. So is Perplexity a wrapper? It answers with other makers' models and searches with its own; the answer depends on which half you count as the product.

How to build an AI wrapper, and what each request costs

How to make a GPT wrapper, or any AI wrapper, comes down to five steps. The third sets the running cost:

  1. Pick one job. One user, one task the general chat box does badly or slowly. What November's small launches charge for, counted by job, is a place to look for one.
  2. Write the prompt and the checks. The system prompt, the inputs you collect and what you verify in the answer before the user sees it.
  3. Choose the model by its bill. Every maker prices input and output tokens separately by the million; the spread is wide (table below).
  4. Build the screen around it. A builder or a coding agent will do; the vibe coding tools and what they cost are compared in our guide.
  5. Charge more than the calls cost. Price per seat, per use or flat; how SaaS products structure their prices is the place to start.

The table prices 12 models from six makers as their own pages printed them on 19 Nov 2026. Anthropic's docs put the unit plainly: “$5 / MTok is $5 for every million tokens.” The last column is the bill for Hummingdraft, a made-up AI email writer, for 1,000 requests of 2,000 input tokens and 500 output tokens each.

AI API prices per million tokens, and what 1,000 requests cost, 19 Nov 2026

AI API prices per million tokens, and what 1,000 requests cost, 19 Nov 2026
ModelMakerInputOutputHummingdraft: 1,000 requestsThe page also says
gpt-6.2-astraOpenAI$8.00$40.00$36.00
gpt-6.2-solOpenAI$2.00$10.00$9.00
gpt-6-lunaOpenAI$0.10$0.50$0.45
Claude Opus 5.5Anthropic$4.00$20.00$18.00
Claude Sonnet 5.5Anthropic$2.00$10.00$9.00
Claude Haiku 5.5Anthropic$0.10$0.50$0.45for prompts up to 100,000 tokens
Gemini 4 CarbonGoogle$2.00 to 30 Nov 2026; $4.00 from 1 Dec 2026$10.00 to 30 Nov 2026; $20.00 from 1 Dec 2026$9.00 to 30 Nov 2026; $18.00 from 1 Dec 2026Google's page prints both prices with these dates
Gemini 3.5 Flash-LiteGoogle$0.30$2.50$1.85
Grok 4.7xAI$2.00$6.00$7.00for prompts under 200,000 tokens
deepseek-v4-proDeepSeek$1.32$3.96$4.62peak hours; off-peak rates are half
deepseek-flashDeepSeek$0.30$1.20$1.20peak hours; off-peak rates are half
Mistral LargeMistral AI$0.50$1.50$1.75

Each price is the maker's own, read on 19 November 2026 between 21:14 and 21:17 UTC: OpenAI, the API docs' Pricing page (Standard, short context, prices per 1M tokens) (developers.openai.com/api/docs/pricing); Anthropic, the Claude docs' Model pricing table (base input tokens and output tokens) (platform.claude.com/docs/en/about-claude/pricing); Google, the Gemini Developer API pricing page (Paid Tier, per 1M tokens in USD) (aistudio.google.com/docs/pricing, which prints by script and was read in a headless browser); xAI, the Models & Pricing docs page (docs.x.ai/developers/models); DeepSeek, the Models & Pricing docs page (cache-miss input, peak hours) (api-docs.deepseek.com/quick_start/pricing); Mistral AI, the pricing page's FAQ, which prints Mistral Large as its example (mistral.ai/pricing). Prices are in US dollars per million tokens, standard rates, before caching and batch discounts, written to two decimals. The last column is our arithmetic for Hummingdraft, a made-up product: 1,000 requests of 2,000 input and 500 output tokens each. Tokens are counted by each maker's own tokenizer, so the same text is not the same number of tokens everywhere. The order is the table's, not a ranking, and naming a model endorses none.

The same 1,000 requests cost $0.45 on Claude Haiku 5.5 or gpt-6-luna and $36.00 on gpt-6.2-astra, 80 times more. At the prices read on 19 Nov 2026, Claude Sonnet 5.5, gpt-6.2-sol and Gemini 4 Carbon land on the same $9.00. Google's page prints two input prices for Carbon, “$2.00 through November 30, 2026.” and “$4.00 starting December 1, 2026.” Its output price moves on the same day, from $10.00 to $20.00, so from 1 December 2026 the same 1,000 requests on Carbon cost $18.00, level with Claude Opus 5.5. DeepSeek's docs add: “Off-peak rates are half of the peak rates.”

Now give Hummingdraft a price: $12 a month, and a user who sends 300 requests a month. On Claude Haiku 5.5 that user costs $0.135 in model calls; on Claude Sonnet 5.5, $2.70; on Claude Opus 5.5, $5.40, or 45% of the subscription; on gpt-6.2-astra, $10.80, or 90%. The model you pick decides most of a wrapper's margin before hosting, payments or support. Prompt caching, which OpenAI, Anthropic, Google, DeepSeek and Mistral AI each price below fresh input, lowers the bill further, and the AI credits startups can apply for can cover the first months of it.

Is an AI wrapper a good business?

It can be, and the census shows why the question comes up: more than half of new AI tools name the same few makers, OpenAI on 146 of 347 pages. Whatever model a wrapper runs on, anyone can rent it too, so it cannot be the moat. Three things a founder controls can be:

  • Data the model does not have. Your customers' records, a corpus you built, the results of every past request. A model maker can match your prompt overnight; it cannot match data nobody else holds.
  • A workflow that fits one job. The steps before and after the model call: what you collect, what you check, where the answer goes next. A ChatGPT wrapper that saves five copy-pastes a day for one profession can be worth more to that profession than the chat box.
  • Distribution. The buyers who already know where to find you. Two products on the same model often differ most in who has heard of them.

Some directories check for something behind the prompt before they list anything. amaze.directory, an AI tool directory in our catalog (our guide to it), sets out “The bar: what we look for” on its submit page, and the second line reads “A real product, not a wrapper around a prompt”. A wrapper with something real behind the prompt passes that bar; a wrapper that is the prompt does not. If you are building one alone, our guide for one-person companies covers the rest of the work.

Where to list an AI wrapper

Distribution starts with being findable, and AI tool directories are where new AI products get listed: all three directories of the census are in our catalog, among 59 AI tool directories from DR 2 to DR 84, 11 of them at DR 50 or above. Every AI directory in our catalog is listed with its Domain Rating, link type and badge rule. Each listing is a page on a site people browse for AI tools, with a link back to your product from a domain search engines already visit. No ads, no outreach: the work is the listings.

Directories also sell speed. AI Tools Up's submit page (our guide to it) offers a free listing, “Listed within 30 days”, and a Fast track card at “$19 (One time)” that reads “No queue, listed within 24 hours”. That is one payment for one place in one queue. For $29, Submitator files 40 directories you choose.

What happens after you pay: the dashboard shows every directory in the catalog with its Domain Rating, and you tick the ones you want. A directory reaches that list once one of our submissions to it has gone all the way through. For each pick we write a description fitted to that site and complete its form, and any CAPTCHA along the way is ours to handle, not yours. Directories that want a badge get it from a single component on your site that carries all of them; we keep each one live, and a new pick never means a new commit. When a listing goes up, its page goes through IndexMachine on our side, the free indexing ($25 value) that comes with every plan. You configure nothing, and the indexing decision itself stays with Google.

Every plan is a single payment: Starter buys 40 directories for $29, Growth 70 for $49 and Scale 110 for $79; All Directories ($149) takes in every directory in the catalog, plus each one added later. Most directories are delivered in 24 to 48 hours, and every directory within 21 days at most. If a directory rejects a submission, by default you get one more directory to pick (All Directories excepted, since it includes them all). A refund can be requested within 30 days of purchase, on the conditions in our terms. 580+ startups have gone through Submitator; how our submissions land is counted in our submission data, and the full catalog lists all 202 sites we submit to, AI and otherwise.

Sources and scope. Each page cited here was fetched with plain GET requests on 19 November 2026, and every quote keeps its author's spelling and punctuation. The census covers 358 tools from DangAI, Turbo0 and The AI Tools Dir under the rules printed beneath its table, fixed before the first home page was fetched. A home page shows what a tool says about models, not what it runs, and no tool from a directory appears here by name. API prices come from the makers' own pricing pages, listed under the price table; Mistral AI's is the example its pricing FAQ gives, and Gemini 4 Carbon carries both prices Google prints, each with its date. ChatGPT: its pricing page, chatgpt.com/pricing. Perplexity: its Pro page, which prints by script and was read in a headless browser, plus two pages of its API docs (Agent API models and embeddings). The harness line: the sidebar of DeepSeek's API docs. AI Tools Up's three cards, read at 21:14 UTC: "Listed within 30 days" (free), "$19 (One time)" with "No queue, listed within 24 hours" (Fast track) and "$249(One time)" (Promoted). AI Tools Up sits in our catalog as an AI tool directory at DR 20, with a guide; the comparison above sets one payment against another, and no plan of ours buys its fast track. amaze.directory: its submit page. Catalog counts and DRs: our export of 17 November 2026, intercepts left out (59 AI tool directories, 202 sites overall). About Submitator: the description we write per directory, the forms and CAPTCHAs we handle, the single badge component kept live, the IndexMachine push, delivery of most directories in 24 to 48 hours and the end-to-end submission of ours that a directory needs before it joins the catalog come from the owner; plan prices and sizes, the indexing bonus and 580+ from the pricing cards and the home page; the 21-day limit, the replacement pick after a rejection (not on All Directories) and the 30 days to request a refund from the terms. No promise of rankings, traffic, signups, revenue, indexing or AI mentions is made here.

FAQ

Is ChatGPT a wrapper?

ChatGPT is OpenAI's own app on OpenAI's own models. Its pricing page, read on 19 Nov 2026, lists GPT-6.1 Sol, GPT-6 Astra, GPT-6 Sol and GPT-6 Luna in its plan table, and the same model family sits in OpenAI's API price list, where any developer can call it. A product that calls those models through the API and adds its own screens is the thing people call a GPT wrapper; ChatGPT is the maker's own front end.

Is Perplexity a wrapper?

Perplexity runs other makers' models and builds its own search around them. Its Pro page lists GPT-6.2 Sol, Gemini 4 Neon, Claude Sonnet 5.5, Kimi K3, GLM 5.3 and Grok 4.7 and says “Perplexity orchestrates the best AI model for each task.” Its API docs describe “The embedding models behind Perplexity Search”, which are its own. Whether that counts as a wrapper depends on whether you see the search or the answer model as the product.

What is the difference between an AI wrapper and an AI agent?

A wrapper sends a request to a model and shows you the answer: one prompt, one reply, whatever interface sits around it. An agent runs the model in a loop with tools, so it can search, call an API, read the result and decide the next step before it replies. Most agents are built on a model someone else trained, so an agent can be a wrapper too; the difference is the loop and the tools.

How much does it cost to run an AI wrapper?

It depends almost entirely on the model you call. For Hummingdraft, a made-up email writer whose requests send 2,000 tokens and get 500 back, 1,000 requests cost $0.45 on Claude Haiku 5.5 or gpt-6-luna, $9.00 on Claude Sonnet 5.5 or gpt-6.2-sol and $36.00 on gpt-6.2-astra, at the makers' list prices read on 19 Nov 2026. The same work can cost 80 times more on one model than on another.

Can an AI wrapper be a real business?

Yes, when the part you own is worth paying for. Anyone can call the same model, so the model is not the moat; the moat is what you add around it: data the model does not have, a workflow that fits one job better than a chat box and a way to reach the buyers who have that job. A wrapper with none of the three competes with the model maker's own app.

How many new AI tools name the model they use?

A little over half, on their home pages. Of 347 new tools on DangAI, Turbo0 and The AI Tools Dir read on 19 Nov 2026, 180 named a third-party model or model maker and 128 named two or more. In a random reading of 30 of those 180, 11 named the model as what the tool runs on; the rest named one they connect to or mentioned one in passing.

Keep reading

Their $19: one fast track. Our $29: 40 directories.

AI Tools Up charges $19 to list one AI tool within 24 hours instead of 30 days. For $29 Submitator files 40 directories of your choosing, or 110 for $79, paid once, with a description written for each site and every badge kept live.