
Photo by NASA Goddard Photo and Video, Public domain, via Wikimedia Commons
An AI tool is a "wrapper" when its intelligence is rented — a call to a model like GPT or Claude through an API — and its own contribution is a text box, a system prompt, and a bill. You can usually tell in under an hour: run the tool's core task through ChatGPT or Claude yourself and see how big the gap is; read the vendor's privacy paperwork for the model providers it names; probe its error messages and refusals for leaked model-provider language; then compare the subscription price against published per-token model prices. Whether it's worth paying has nothing to do with the label. It depends entirely on what sits on top of the model — your data, live integrations, workflow controls, review steps, support — and whether that layer saves you more than it costs. Plenty of wrappers are good buys. What isn't a good buy is paying a premium for a prompt you could write yourself.
What "wrapper" actually means (and what it doesn't)
Almost every AI feature in business software calls a hosted model. OpenAI sells API access to its models; Anthropic does the same; and the same models increasingly reach buyers through third-party clouds — OpenAI models are now available on Google Cloud's Vertex AI platform, billed to the customer's Google Cloud account and run inside Google's security perimeter (OpenAI on Google Cloud). In other words, "built on GPT" is plumbing, not a scandal.
The useful distinction isn't wrapper vs. non-wrapper. It's:
- Thin wrapper: a UI, a prompt, and an API call. Remove the model call and the product stops working.
- Thick wrapper: the model is one component inside a system the vendor owns — connectors to your CRM or ticketing stack, a proprietary data pipeline or retrieval index, permissioning and audit logs, evaluation and quality checks, human review, and support.
A thin wrapper is a subscription to an interface. A thick wrapper is software you'd otherwise have to build. Your job as a buyer is to figure out which one you're being quoted.
Test 1: The swap test
This is the single most informative thing you can do, and it needs no technical background.
- Pick one real job the tool claims to do — a listing description, a support reply triage pass, a first-draft contract summary.
- Give the tool your actual input, and save the output.
- Open a general assistant you already have (ChatGPT is free, with paid subscription plans available, and Claude has a free tier too), describe the same task in your own words, and paste the same context. If you want to give the free assistant a fair shot, build a reusable custom GPT or project with the tool's instructions rather than winging a single prompt. And compare across providers rather than trusting one: our ChatGPT vs Claude breakdown shows how differently the same kind of request can land.
- Compare. If the outputs are interchangeable, the product largely is the prompt. If the assistant flails because it can't reach your files, your records, your codebase, or live data, the product is doing work the model alone can't.
Then push one step further with a second question: would this product still exist if its model disappeared tomorrow, or if the model maker shipped the same feature natively? A durable product keeps its value because it owns the data, the integrations, and the workflow. A thin wrapper collapses, because the model was the product.
Test 2: Read the paperwork the vendor is legally required to publish
Vendors that process customer data through third parties generally publish a subprocessor list in their privacy or DPA documentation. Serious vendors list model providers plainly. Datadog's published subprocessor list, for example, names Anthropic, OpenAI, Fireworks.ai, Baseten, Nebius, Databricks, and Together Computer under "AI services" (Datadog subprocessors). That's what disclosure looks like.
So search the vendor's site for "subprocessor", "model provider", "large language model", "third-party AI provider", and "OpenAI", "Anthropic", "Google Cloud", "Bedrock". What you learn:
- Named providers plus a data-handling story — where inference runs, retention period, whether your content trains models, whether they support Zero Data Retention or equivalent terms — is a vendor that understands its own stack.
- "Proprietary AI" in the marketing, one model provider in the subprocessor list is a contradiction worth asking about.
- No answer at all is the biggest red flag of all. If a vendor can't tell you where your data is processed, that's not an oversight.
Do your own homework on the provider side so you can ask sharp questions. OpenAI's API documentation states that data sent to the API is not used to train or improve its models, that abuse-monitoring logs are retained up to 30 days by default, and that Zero Data Retention requires prior approval rather than a request flag like store: false (OpenAI: Your data). For consumer ChatGPT, by contrast, OpenAI says content may be used for training unless you opt out, while content from business offerings is not. Claude's plan page lists "no model training on your content by default" for Team and Enterprise (Claude pricing).
The practical point: paying for a wrapper does not automatically mean better privacy. A reskinner of a consumer chatbot may give you weaker data terms than the provider's own business plan.
Test 3: Probe the failures
Wrappers show their seams where things break. During a trial, spend twenty minutes trying to make the tool fail:
- Paste an enormous document. Error messages that mention "tokens", "context length", or "context window" are model-provider vocabulary leaking through.
- Look at raw error strings. Identifiers beginning with a model family name settle the question outright.
- Ask for something the model makers restrict and compare the refusal wording with ChatGPT's or Claude's. Near-identical refusals, or a message that cites a provider's usage policies, point to a pass-through call.
- Watch outages. If the tool degrades whenever a model provider's status page reports an incident, it depends on that provider — ask what the vendor's fallback is.
- Check which features are missing. The same model behaves differently depending on how it's delivered. OpenAI's own comparison shows hosted web search, hosted file search, hosted code execution, image generation, batch processing, and the Chat Completions endpoint are available on the first-party API but not through Google Cloud (feature availability table). A tool that can't browse, generate images, or run code may simply be calling a model through a channel that doesn't offer those features — not having "custom AI".
Test 4: Do the model math
This is where price objections get specific. Published API list prices are per million tokens, input then output. OpenAI's pricing page lists, for standard short-context requests: gpt-6-luna at $0.10 / $0.50, gpt-6.1-terra at $1.00 / $5.00, gpt-6.1-sol at $2.00 / $10.00, and gpt-6-astra at $10.00 / $50.00, with Batch pricing roughly half those rates (OpenAI API pricing). Anthropic lists Haiku 5.5 at $0.10 / $0.50 for prompts under 100K tokens, Sonnet 5.5 at $2 / $10, and Opus 5.5 at $4 / $20, also with about 50% off via batch processing (Claude pricing).
Worked example, using round numbers you can swap for your own volume. Say a tool drafts 1,000 short documents a month, each sending roughly 2,500 input tokens and returning 700 output tokens. That's 2.5M input and 0.7M output tokens.
| Model tier | Input cost | Output cost | Monthly model cost |
|---|---|---|---|
| Haiku 5.5 / gpt-6-luna ($0.10 / $0.50) | $0.25 | $0.35 | ~$0.60 |
| gpt-6.1-terra ($1 / $5) | $2.50 | $3.50 | ~$6.00 |
| Sonnet 5.5 / gpt-6.1-sol ($2 / $10) | $5.00 | $7.00 | ~$12.00 |
| Opus 5.5 ($4 / $20) | $10.00 | $14.00 | ~$24.00 |
For a solo user, raw model compute for that workload is single-digit dollars. If the vendor charges $40 or $90 a month, the difference is not the model — it's connectors, data pipeline, compliance, support, and margin. That's fine if those things exist and you use them. It's a bad deal if the difference is a nicer text box.
Also compare against the provider's own plans rather than only against API calls — including the free plans and paid tiers of ChatGPT that many teams already pay for without realising. Claude's Team plan is $20 per seat per month billed annually ($25 monthly), with Enterprise starting at $20 per seat per month plus usage at API rates on a 20-seat minimum (Claude pricing). If a third-party tool costs several times that per seat and produces interchangeable output, the burden of proof is on the vendor.
When the wrapper is worth it
Look for these, and buy when several are true:
- It touches systems you already use. Reading your CRM, cross-referencing tickets, writing back, notifying Slack. Integration is where most of the real work lives, and it's the part a general assistant can't do.
- It has controls a chat app lacks: approvals, audit logs, per-user permissions, usage caps, retention settings.
- It knows things the base model doesn't: your documents, a licensed corpus, live prices, case histories, corrections your team has already made.
- It gets better by month six for reasons other than the underlying model being upgraded.
- The vendor names its models and answers data questions in writing, including what happens on model deprecation.
- The price tracks the hours it saves you, not the tokens it spends.
When it isn't — and what to do about it
Walk away, or renegotiate, when the output matches a general assistant almost word-for-word with no integrations to justify the price; when "proprietary AI" sits next to a single model provider in the paperwork; when there's no answer on training, retention, or model changes; and when the monthly fee is many multiples of the model cost for work your team could replicate in a shared project in an afternoon.
If you're already on an annual contract, three moves limit the damage: export your data and prompt templates now while you have access; get written terms on model swaps, price changes, and exit; and keep a parallel workflow in your own assistant so renewal is a choice, not a hostage situation.
FAQ
Is every tool that uses GPT a rip-off? No. Renting a model through an API is a normal product decision, and OpenAI explicitly documents developers building products on its models. The problem is pricing that assumes proprietary technology when the vendor only owns a prompt.
Why not just use ChatGPT or Claude directly? Because you'd be the integrator. If your job is drafting text you'll paste somewhere, a general assistant is usually enough — which is why our AI tools for small business guide leans on a mix of general assistants and workflow tools rather than single-purpose AI apps. If your job requires the tool to read your records, act on them, and keep humans in the approval loop, that integration work is real and has real cost.
Does a wrapper mean my data trains someone's model? Not necessarily. OpenAI states API data isn't used for training or improvement absent an opt-in; OpenAI says content from ChatGPT Business and Enterprise offerings isn't used for training; Claude lists no model training on content by default for Team and Enterprise. What matters is which plan the vendor is on and what it commits to contractually — ask in writing.
Will a vendor tell me which model it uses? Sometimes, and the willingness is itself a signal. Vendors comfortable with their architecture will name providers and data terms. Vendors who answer "our proprietary AI engine" to every variation of the question are telling you something.
If it's a thin wrapper, is the money always wasted? Not always. A cheap wrapper that genuinely saves you an hour a week is a fine purchase — buy it monthly, keep your own copies of the prompts and examples you feed it, and re-evaluate when the model makers inevitably ship that feature as a native button.
Sources
- OpenAI API pricing documentation — https://developers.openai.com/api/docs/pricing
- OpenAI, "Your data" (API training and retention terms) — https://developers.openai.com/api/docs/guides/your-data
- OpenAI on Google Cloud (feature availability, retention) — https://developers.openai.com/api/docs/guides/google-cloud
- OpenAI Help Center, "What is ChatGPT?" — https://help.openai.com/en/articles/6783457-what-is-chatgpt
- Claude pricing and plan comparison — https://claude.com/pricing
- Datadog third-party subprocessor list — https://www.datadoghq.com/subprocessors/
Disclosure: this article may contain affiliate links.