Cost & Models
How ArkAssist chooses an AI model, how usage is metered and attributed, and how cost differs between a single structured action and a multi-step Agent run.
ArkAssist is powered by AI models, with more capable options unlocked on higher plans. This page explains how a model is chosen, how usage is metered, and where cost comes from.
How a model is chosen
For each ArkAssist feature, the model used is decided in this order:
- Your saved preference — a per-feature setting, then your general preferred model.
- An explicit choice — the model you pick in the model selector for that request (subject to your plan tier).
- The feature default — a sensible default for each feature (chat, section writing, outline, compliance, budget, success prediction).
- A general default — a fallback default when no preference or choice applies.
Your subscription tier governs which models you may use. Asking for a model above your tier shows an upgrade prompt rather than silently downgrading; a saved preference that is no longer available on your plan is simply skipped in favour of the next option.
Metering and attribution
Every AI exchange is recorded for usage limits and transparency:
- Each message records the model you requested versus the model that actually answered, the amount of work it required, and an estimated per-message cost based on the model that served the request.
- If a different model answers than the one you requested, that difference is recorded so it is visible to you rather than hidden.
Whichever door you use — the structured workspace or the ResearchArk Agent — usage is metered the same way, because both perform the same underlying work.
Single action vs. multi-step run
The two doors have different cost shapes for the same outcome:
- A structured-workspace action (generate one section, run one compliance check) is a single AI request — predictable, bounded cost.
- An Agent run that drafts a whole proposal combines several requests (create → outline → sections → budget → compliance). It is faster and more guided, but its total cost is the sum of those steps.
If you are watching spend, prefer single structured actions for targeted work and reserve full Agent runs for when you want an end-to-end first draft.
Reliability
If an AI request runs into a temporary problem, ArkAssist retries it automatically a limited number of times, and a stalled streaming response surfaces an error you can retry rather than hanging indefinitely. Retries are limited and accounted for, so they cannot silently inflate your cost.
Application Management
Track grant applications through their full lifecycle with search, filtering, section-based writing, AI content generation, compliance checking, and team collaboration.
ArkSphere
Research networking module for discovering collaborators, building research communities, visualizing professional networks, and browsing EC-funded organizations and people.