LLM models and APIs
Verboo, Venice AI and Drael vs NotHoly for security work
Verboo, Venice AI and Drael next to NotHoly, November 2026: billing, API format, licence terms, model bases and where tools run, with sources.
Drael vs NotHoly, November 2026: an unnamed operator, one hidden model, a US$ 70 API entry and code that runs on Drael's server. Our verdict.
Drael (drael.sh) sells a subscription to one model it will not identify, from an operator its terms never name, with tools that run on its own server. NotHoly sells prepaid access to models with published bases, from a company named in its terms, and its infrastructure never touches a target. For professional security work we do not consider Drael an option.
Data collected on 9 and 10 November 2026 from Drael's site, its documentation and legal pages, and from NotHoly pricing. Prices change, so check both before you buy. We did not open a Drael account, and nothing here rates the quality of its answers.
Every row below comes from a page you can open. The ones that decide a client job are the seller, the model and where the code runs, and Drael loses all three.
| Drael (drael.sh) | NotHoly | |
|---|---|---|
| Seller named in the terms | No, only "the operator" | Yes, with CNPJ, governing law and venue |
| Models | One alias, drael-v1, source weights undisclosed |
A catalogue, with the base of each model published |
| Lowest entry to the API | R$ 199 / US$ 70 for four weeks | US$ 5 top-up, credits do not expire |
| Pricing model | Subscription block; extra usage metered | Prepaid, metered per token; rate per model on pricing |
| Output control | Reasoning forced on through a key, max_tokens dropped |
You set max_tokens and reasoning; the gateway passes them through |
| Context | 128K on Pro, 248K on Max | Listed per model on models |
| API format | OpenAI-compatible | OpenAI-compatible and Anthropic-compatible |
| Code execution | On Drael's server, network on by a flag | On your machine through the CLI, with your approval |
| Teams | Nothing in the public documentation | Organizations, roles, workspaces and an audit log |
| Payment | Docs list card, Pix and crypto; the public endpoint returned one crypto processor | Card and crypto, no Pix |
| Vulnerability reporting | security.txt returns 404 |
Published security.txt and security page |
Drael's published input meter reads low, but you reach it only after the US$ 70 block, and its output line, the expensive one, is billed with reasoning you cannot switch off through a key.
Drael's terms call the seller "the operator" and give no company, registration, address, governing law or venue. The only contacts are a support mailbox and a Discord server. NotHoly's terms name the holder by CNPJ, with governing law and venue.
The same terms put third-party claims on you: "If someone makes a claim against the operator because of how you used Drael or what you did with an answer, you take responsibility for it, including the reasonable cost of defending it." Drael's own liability stops at "what you paid in the three months before the claim". So you indemnify a party you cannot name, and that party caps its own exposure at three months of subscription. Our reading of the terms goes clause by clause.
The chat documentation describes drael-v1 as "a public alias over weights that are ours to change". It does not say what those weights are. An unknown model ID is answered by the same alias with no error, and the limits page says "we prepend a system prompt you never see".
The documentation also records how the model misbehaves. With reasoning on and max_tokens at 4000 it "returned an empty answer in 22 of 24 runs". With reasoning off, "a terminal session ran one command thirty times". Drael's answer was to force reasoning on every API call and discard your max_tokens. The drael-v1 post covers what that does to a finding you must reproduce.
NotHoly publishes each base on the models page. Most are abliterated, meaning the refusal direction was removed from the weights and the rest kept, as our explainer describes. We did not train them, and they can still err or refuse.
The Drael API exists only on the four-week block or the annual plan, so the smallest API purchase is R$ 199 / US$ 70. A weekly plan answers plan_required. Drael's "month" is four weeks, which makes 13 payments a year. NotHoly starts at a US$ 5 top-up, and credit does not expire.
Drael drael-v1 public meter, per million tokens, USD |
Input | Cached input | Output |
|---|---|---|---|
| Extra usage and Enterprise | 0.18 | 0.04 | 1.40 |
Inside the plan Drael promises volume "within fair use" and does not publish the per-minute rate, because it "is being set against real usage". NotHoly meters per token with no plan around it, each model's rate and base on pricing and models. The pricing post and the API cost post have the arithmetic.
Drael's run_code, web_fetch and breach_lookup execute on its side, in a loop of up to ten turns, and the tools documentation says "You never see the intermediate turns". Its privacy page says "A page you ask it to fetch sees a request from our server". On a client job, that puts an address nobody listed in the scope on the client's logs. Our post on server-side execution explains why we call that a flaw.
The same documentation describes the capacity behind it: "Everybody being served at once shares one machine", and a model stuck in a tool loop "on a GPU this product owns is an outage rather than a bill". Drael publishes no capacity figure and no SLA. In our opinion those sentences describe a service running on very little hardware.
On NotHoly no server tool scans, exploits or sends active traffic to a third party. The CLI runs commands on your machine, from your address, after you approve each one.
The nav link "Pricing" is an anchor on the home page, and /pricing returns 404 while the terms tell you to check "the plan page". The footer "Help" link redirects to sign-in. The quickstart samples call https://YOUR-HOST/v1. The site is offered in nine languages and the legal pages exist in two. Our view is that this is a product shipped before anyone reviewed it, and the post on Drael's site lists what we found.
Drael fits a solo user who wants an uncensored chatbot with image generation and Pix and will never name the tool in a contract or a scope. NotHoly fits work with a client and a scope: a seller named by CNPJ, a model whose base you can cite, commands that run from your own address, and a team account with roles and an audit log. NotHoly holds no SOC 2 or ISO 27001 and requires an email, listed on the trust page; it makes no images and takes no Pix.
The site of Drael, a subscription with uncensored chat, one model called drael-v1, server-side code execution, image generation and breach lookups. Its terms identify the seller only as "the operator".
We would not put client work on it. The seller is unnamed, the model is undisclosed, traffic leaves from Drael's server and the terms move third-party claims to you.
Yes, OpenAI-compatible, on the four-week or annual plans only. The lowest entry is R$ 199 / US$ 70.
One that executes on your machine, names its model and names itself in the contract. NotHoly does all three; see pentesting. We also compared Venice AI with NotHoly.
Drael is an unnamed seller with one unnamed model and a server that acts in your place. That fits someone who wants an uncensored chatbot with image generation and Pix, neither of which NotHoly offers. It does not fit work with a scope and a client. NotHoly holds no SOC 2 or ISO 27001 and requires an email; the trust page says what we do hold. Start with US$ 5 on NotHoly pricing.
LLM models and APIs
Verboo, Venice AI and Drael next to NotHoly, November 2026: billing, API format, licence terms, model bases and where tools run, with sources.
LLM models and APIs
Flat unlimited AI plans usually limit speed, not tokens. How Verboo, Venice.ai, Drael and prepaid per-token credit compare on terms and bills.
LLM models and APIs
Moving an OpenAI-compatible workflow from Verboo, Venice AI or Drael to NotHoly: base URL, key, model IDs, billing, licence and tools.