drael-v1: one hidden model, and docs that list its bugs

Drael will not say what model it uses. drael-v1 is, in its own words, "a public alias over weights that are ours to change". It runs behind a system prompt you never see, ignores your max_tokens, and answers to any model name you send. Its documentation also records the model failing in ways we would not accept in a tool for client work.

Data collected on 9 and 10 November 2026 from Drael's quickstart, chat documentation and limits page, and from NotHoly pricing. We did not call the Drael API. Every statement about the model's behaviour below is a quote from Drael's documentation.

#What model does Drael use

Nobody outside Drael knows. The quickstart says "The model id is drael-v1, and it is the only one". The chat documentation adds that it is "a public alias over weights that are ours to change". No base model, publisher, parameter count or licence is given.

Question	What Drael's docs answer
Which base model?	Not stated
Can the weights change?	Yes: "weights that are ours to change"
Is a change announced?	Nothing in the documentation says so
What if I send another model ID?	"an id nobody serves is answered on it rather than refused"
Is there a system prompt?	"we prepend a system prompt you never see"
How many models were there before?	"There were two", and the second is "gone"

#The failures Drael's own docs record

The chat documentation explains its design decisions by describing what went wrong. Read as a list, it is a bug log.

With reasoning on and max_tokens at 4000, the model "returned an empty answer in 22 of 24 runs, and at 8000 in 12 of 24", each time with no error and no content. Drael's fix was to drop max_tokens on every call through a key.

With reasoning off, "the model re-does its last step in a tool loop: a terminal session ran one command thirty times (19 October 2026)". The default reasoning setting "was none for one evening and agent clients repeated the same completion on a third of their turns". Drael's fix was to force reasoning on for every API call.

The tools documentation adds that "the previous product reached 13.218 tokens of tool schema and paid for it in invented tool names and shallower answers".

Publishing this is unusual, and we take the numbers at face value. What they describe is a model that returns nothing under an ordinary output limit and repeats commands in a loop unless it is made to reason, patched by removing the controls a developer would use to manage it. On a client engagement, a tool that reruns the same command thirty times is a problem you have to explain to the client.

#What you cannot control

Parameter	What Drael does with it
model	Any unknown ID is answered by drael-v1, with no error
max_tokens	"through a key the ceiling is always dropped"
reasoning_effort: none	Answered at low; "There is no off through a key"
priority	"discarded"
system	Merged with a prompt you never see
The date	Repeated into your last message

A typo in a model name never surfaces, because the alias answers anyway. A cost ceiling per call does not exist. Part of the prompt is Drael's and is not shown to you. Each of these is a small loss of control, and together they mean you cannot state what configuration produced a given answer.

#Why this ruins a finding

A finding in a report has to be reproducible: same input, same tool, same version. With drael-v1 you cannot name the version, because there is none. If a client retests in three months and the model behaves differently, you have no way to say whether the weights changed, the hidden prompt changed, or both.

Evaluation has the same problem. A refusal rate or an accuracy score measured this week describes an alias that may point somewhere else next week. The HackerOne Hacker-Powered Security Report 2026 counts a 557% rise in valid system prompt leak reports in a year, which is a reminder that hidden prompts are both common and fragile.

#What NotHoly publishes

Every NotHoly model names its base and its context window on the models page, and most are abliterated. The base is a model the publisher released, so you can write it, with the date, into a report.

Abliterated means the refusal direction was removed from the weights and the rest kept; our explainer and abliterated vs uncensored models cover the technique. Arditi et al. found refusal mediated by a single direction across 13 open models, and the Heretic README reports the trade between refusals and divergence from the original. We did not train these weights and do not own them. They can still be wrong, and they can still refuse.

#Where each fits

drael-v1 fits a user who just wants an answer and never has to reproduce it: the model, the hidden prompt and the limits are Drael's to change. NotHoly fits work that has to be repeatable and citable, where each model names its base and its context on the models page, you set your own max_tokens and reasoning, and you can write the base model and the date into a report. The weights are not ours; we did not train them, and they can still err or refuse.

#Frequently asked questions

#What is drael-v1?

The only model ID on Drael's API. It is an alias, and Drael reserves the right to change the weights behind it.

#What model does Drael use?

Drael does not disclose it.

#Does Drael respect max_tokens?

Not through an API key. The documentation says the ceiling "is always dropped" there.

#Can I turn off reasoning on the Drael API?

No. none and off are answered at low.

#Why did a wrong model name still work on Drael?

Because any ID "nobody serves" is answered by drael-v1 without an error.

#Which model should I cite in a report that used NotHoly?

The model name and its published base from the models page, with the date.

#Verdict

drael-v1 is a name with nothing behind it that you can check. The weights are secret and replaceable, the prompt is partly hidden, the output limit is ignored, and the documentation describes loops and empty answers as known behaviour. We would not sign a report built on it. NotHoly names every base and takes a US$ 5 top-up on NotHoly pricing; the full comparison and the API cost post have the rest.