Safeguard models

Griffin, Eagle, Lion and Zero in the model catalog: what each is for, their parameters, context windows and prices, where they run, why Zero uses no model tokens, how their data is handled, and their status through the API.

Safeguard builds its own security models. The catalog at safeguard.sh/models lists them next to the third-party models: three models (Griffin, Eagle and Lion) and Zero, which answers data questions with no model at all.

None of Safeguard's own models can be called through the inference API yet. GET /models?all=true lists them with their current status.

The lineup

Model	Model id	What it is for	Parameters	Context window	Catalog price per million tokens
Griffin	`safeguard/griffin`	Remediation and reasoning: reachability, patch authoring and fix pull requests.	750 billion	1.1 million tokens (1,100,000)	$0.0001 input, $0.0001 output
Eagle	`safeguard/eagle`	Discovery and triage: malicious package and typosquat classification, and path ranking.	70 billion	1 million tokens (1,000,000)	$0.00001 input, $0.00001 output
Lion	`safeguard/lion`	Compliance and narrative work: evidence mapping, audit summaries and policy drafting.	15 billion	128K tokens (128,000)	$0.000001 input, $0.000001 output
Zero	`safeguard/zero`	Data questions with no model at all: it reads a question into a deterministic, replayable query.	None: Zero has no model weights	Not applicable	$0. Zero uses no model tokens.

Griffin, Eagle and Lion are one model each. None of them comes in sizes: there is no Lite, S, M, L or XL.

Each has a page in the catalog: Griffin, Eagle, Lion and Zero.

Zero uses no model tokens

Zero has no model weights. It turns a question about your data into a deterministic query that can be replayed, and runs that query in process on the Safeguard platform. No language model is called and no model tokens are consumed, which is why its catalog price is $0 per million tokens.

Where they run

On the Safeguard platform, all four can run on-prem, air-gapped, in a private cloud (VPC) and in public cloud. See Deployment options.

They are available in 245 countries and territories (the ISO 3166-1 list minus the exclusions). They are not offered in Cuba, Iran, North Korea or Syria, nor in the Crimea region of Ukraine or the so-called Donetsk People's Republic (DNR) and Luhansk People's Republic (LNR) regions of Ukraine.

Through the API Platform, none of them can be called yet, in any of these options.

How their data is handled

The catalog's data policy for these models says:

	Griffin, Eagle and Lion	Zero
Where prompts go	Safeguard, on its own infrastructure	Nowhere: Zero runs in process on the Safeguard platform
Training	Customer data is never used to train shared models.	No model provider is involved. Customer data is never used to train shared models.
Region	The region your Safeguard workspace runs in	The region your Safeguard workspace runs in

Through the API

Their catalog status is not yet available, so a request naming one is refused with 503 model_unavailable.
Through the API, Safeguard's own models are served only by Safeguard-hosted inference, never by a third-party provider.
Because no third party receives the prompts, the third-party launch checklist does not apply to them. They still need a successful end-to-end test call before they can be called. See Models.
When they become callable, they use the same endpoint, request format and billing as every other model, at the catalog prices above.

In the rest of Safeguard

These models also appear across the Safeguard platform. The Griffin AI and AI Models pages describe them there; the AI Models page calls the compliance model Lino, which is Lion in the catalog.