Finance teams approving multi-year security software spend care about one number above all: how accurately can we predict what this costs next year, and the year after? Platforms whose price has an unbounded usage component — token spend, per-finding charges, dynamic scale-up — produce year-two surprises that wreck budget forecasting. Griffin AI and Mythos-class general-purpose AI-for-security tools have very different predictability shapes, and the difference is rooted in architecture rather than in the specific deal structure.
What makes pricing predictable
Three properties:
- Stable unit of pricing. Per-developer, per-repo, per-service — anchors that don't change unexpectedly.
- Token spend as a bounded component rather than an unbounded one.
- Fixed-price incident spikes — or at least bounded spike behaviour during IR.
All three need to hold for the finance team to forecast accurately.
Where architecture creates predictability
Two direct consequences of the engine-plus-LLM design:
LLM calls are gated, not constant. The deterministic engine handles routine analysis. Griffin AI's LLM calls happen at specific, measurable reasoning points. The rate of LLM calls scales with findings requiring reasoning, not with scan volume. This keeps token spend bounded even as scope grows.
Model tiering prevents price-performance drift. Cheap models handle routine work; expensive models handle high-leverage reasoning. The mix is engineered for cost-effectiveness. An upgrade that moves a task from the cheap tier to the expensive tier is surfaced as a release note, not a silent cost increase.
Where pure-LLM tools lose predictability
Three structural reasons:
Token spend scales with scan volume. Every analysis step is a model call. Adding 10 repos means 10x the model calls.
Model upgrades change cost. A frontier model upgrade that moves the platform to a more capable (and more expensive) model passes the cost through to customers. The customer sees the token-spend line rise without a corresponding change in their own usage.
Incident spikes are unbounded. During an incident, the platform is being queried heavily. Every query is a model call. Spend spikes when budget attention is lowest.
A concrete comparison
Consider a 300-developer organisation deploying an AI-for-security tool with a three-year contract.
With Griffin AI's architecture:
- Year 1: base license + bounded token spend (token spend typically 5-15% of license).
- Year 2: license renewed; token spend grows only if scope grows. Predictable.
- Year 3: same shape.
With a Mythos-class tool scaling tokens per operation:
- Year 1: base license + token spend (initially comparable).
- Year 2: token spend grows even without scope expansion because the model has been upgraded. Add incident-driven spikes.
- Year 3: token spend is now 40% of total cost; license negotiation focuses on capping it.
The difference is not in deal quality. It is in which costs are bounded by architecture and which are unbounded.
How Griffin AI's pricing model reflects this
Customer-facing pricing focuses on:
- Per-developer or per-service pricing for the platform license.
- A bounded token-spend component that is estimated per-customer and forecast for the contract duration.
- Spike behaviour during incident scenarios is documented and scoped.
Customers can build an honest three-year TCO from the contract terms. Mythos-class contracts typically require customer-side modelling to produce an equivalent number, and the modelling is sensitive to assumptions about model-vendor pricing that the customer cannot control.
What to evaluate
Three concrete asks during procurement:
- A committed three-year pricing schedule with token-spend caps or at least clear cost-drivers.
- A documented behaviour during incident scenarios — how does spend change?
- A price-change notification policy — how much notice before a model upgrade that increases cost?
The answers determine whether the vendor's pricing is finance-approvable for multi-year commitment.
How Safeguard Helps
Safeguard's pricing structure reflects the engine-plus-LLM cost architecture. License is the primary line; token spend is bounded and predictable; incident-spike behaviour is documented. Customers building three-year TCO models have the information to do so confidently. For finance teams that have been burned by year-two cost surprises on AI-heavy tooling, predictability is the property that justifies multi-year commitment.