The simplest LLM architecture routes every request to one model. It works; it's expensive. Task-routed architectures match the capability tier of each task to an appropriately-tiered model — small models for routine work, large models for complex reasoning. For security workloads specifically, task routing is the highest-leverage cost-efficiency lever available. Griffin AI is built around this pattern.
What task routing means in practice
Three tiers:
- Small models (Haiku-class, Gemini Flash, small open-weight). Bulk classification, summarisation, routine extraction.
- Mid-tier models (Sonnet-class, Gemini Pro). Multi-step drafting, standard analysis, fix-PR generation.
- Large models (Opus-class, Gemini Ultra). Complex reasoning, novel exploit hypothesis, hard edge cases.
Each task routes to the tier that produces adequate quality at minimum cost.
Where task routing pays off
Three workloads:
- Finding deduplication. Small model; massive volume; routine pattern matching.
- Fix-PR drafting. Mid-tier model; requires structured reasoning but not frontier capability.
- Zero-day hypothesis. Large model; complex multi-hop reasoning on specific taint paths.
A flat architecture would either overpay (running everything on Opus) or under-deliver (running everything on Haiku).
How Griffin AI implements it
Four routing decisions:
- Task type determines tier baseline.
- Complexity signals from the engine output can upgrade to a larger model.
- Confidence thresholds trigger fallback to a larger model when the smaller one's confidence is below the bar.
- Eval-gated fallbacks route tasks known to need frontier capability to the right tier automatically.
The routing logic is itself evaluated in the eval harness.
What task routing does NOT mean
Two common misconceptions:
- Task routing is not "one model per customer." Routing decisions are per-request, not per-deployment.
- Task routing is not a quality compromise. The small-model tiers are used only where quality measures adequate; where it isn't, the request routes up.
How Safeguard Helps
Safeguard's Griffin AI implements task routing across frontier model tiers. Cost efficiency and quality gates are both managed automatically. For customers whose previous AI-for-security tools wasted budget on uniform model usage, task routing is the architectural choice that delivers the same quality at materially lower total cost.