Three ways to run AI. Real costs from real benchmarks.
Claude Desktop (flat subscription), Claude Code (allowance-based), or the built-in API gateway (pay-per-token). Each option connects to the same 9 MCP servers and 254 GRC tools. The cost model is the only difference.
Choose your model
Subscription vs pay-per-token
Claude Desktop
Subscription
Cost
$20/mo Pro · $100/mo Max
MCP servers run as local stdio processes. Claude Desktop spawns them directly. You get flat-rate usage within Anthropic's fair-use limits.
Best for
Daily GRC queries, routine risk checks, quick lookups
Trade-off
Rate-limited during peak usage. No autonomous workflows, no council, no approval UI.
Claude Code
Subscription + Allowance
Cost
$20/mo (Max plan included)
Claude Code reads .mcp.json from the project root and connects to all 9 MCP servers automatically. Usage counted against your monthly allowance.
Best for
Development, exploration, cross-domain queries with all 9 servers
Trade-off
Same rate limits as Claude Desktop. Usage allowance may cap heavy sessions.
Built-in Gateway (API)
Pay-per-token
Cost
No subscription — pay only for what you use
The RiskReady gateway routes queries to domain MCP servers, runs the AI Council, executes scheduled workflows, and manages the approval queue — all via the Anthropic Messages API with your own API key.
Best for
Production use, autonomous workflows, council deliberations, approval-gated mutations
Trade-off
Requires ANTHROPIC_API_KEY. Cost scales with usage and model choice.
API pricing
Per-token rates (Anthropic Messages API)
Real benchmarks
Actual costs from a live GRC database.
Tested against ClearStream Payments: 15 risks, 40 controls, 30 scenarios, 8 KRIs, 12 active incidents. These are real token counts, not estimates.
Single query
“Show me the top risks”
6-agent council (Haiku)
“Comprehensive security posture review”
6-agent council (Opus)
“Comprehensive security posture review”
Monthly estimates
API costs by usage pattern
| Profile | Usage | Haiku | Sonnet | Opus | Desktop |
|---|---|---|---|---|---|
| Light user | 10 queries/day, no council | ~$2/mo | ~$8/mo | ~$40/mo | $20/mo (Pro) |
| Active user | 30 queries/day, 2 councils/week | ~$8/mo | ~$30/mo | ~$150/mo | $20/mo (Pro) |
| Power user | 50 queries/day, daily council, scheduled workflows | ~$15/mo | ~$60/mo | ~$300/mo | $100/mo (Max) |
Recommendation: Start with Claude Desktop Pro ($20/mo) for daily queries. Switch to API with Haiku for autonomous workflows and scheduled runs. Use Opus only for board-level reports and audit preparation.
Built-in savings
How the gateway keeps costs down
Tool search (96% savings)
defer_loading lets Claude discover tools on demand instead of loading all 254 schemas — input drops from 228K to 8K tokens per request.
Prompt caching (90% discount)
System prompts are cached for 5 minutes. Repeated queries within a session get 90% off the cached portion.
Council batching
Council members run in pairs of 2 to stay within memory limits, with per-member token tracking.
Model selection per task
Use Haiku for daily queries ($0.007), Sonnet for analysis, Opus only for board reports and audit prep ($10).
Full benchmark data
Run your own benchmarks against your data.
The gateway logs per-member token usage for every council session. Deploy the demo, connect your API key, and see exactly what your workloads cost.