Skip to content

Cost Estimates

Foundry’s infrastructure cost scales with usage. The platform costs are modest at evaluation and early production tiers; AI API usage is the primary variable cost.

ServicePlanMonthly EstimateNotes
VercelPro~$20Compute offloaded to Convex and Cloudflare. Vercel handles SSR, Edge Middleware, and OAuth callback routes.
Cloudflare WorkersPaid ($5/mo)~$5-15Agent worker is low-volume (structured analysis requests). Sandbox worker costs depend on container usage.
Convex CloudFree / Pro$0-25Free tier covers evaluation. Pro tier ($25/mo) for production with higher function call limits and bandwidth.
ClerkFree / Pro$0-25Free tier covers up to 10,000 MAU. Pro tier at $0.02/MAU beyond that.
Total infrastructure~$25-65/moEvaluation can run on free tiers for all services.

AI costs depend on which features you use and how frequently. Here are estimates per operation based on the context assembly pipeline and model tiers.

OperationModelInput TokensOutput TokensEst. Cost
Document analysisOpus 4.6~8,000~4,000~$0.30
Task decompositionSonnet 4.5 v2~5,000~3,000~$0.04
Subtask generationSonnet 4.5 v2~4,000~2,000~$0.03
Sprint planningSonnet 4.5 v2~6,000~3,000~$0.04
Gate evaluationSonnet 4.5 v2~5,000~2,000~$0.03
Risk assessmentSonnet 4.5 v2~4,000~2,000~$0.03
Health scoring (per workstream)Sonnet 4.5 v2~3,000~1,000~$0.02
Daily digestSonnet 4.5 v2~2,000~500~$0.01
Skill execution (sandbox)Sonnet 4.5~10,000~5,000~$0.05

1 program, 10-20 requirements, light usage

CategoryMonthly Estimate
Document analysis (5 docs)~$1.50
Task decomposition (20 reqs)~$0.80
Subtask generation (20 tasks)~$0.60
Sandbox executions (10 runs)~$0.50
Health scoring + digests~$1.00
AI total~$4-5
Infrastructure$0 (free tiers)
Grand total~$5/mo

The sandbox worker provisions Docker containers on Cloudflare. Container costs are separate from the Workers Paid plan.

  • Compute: Billed per GB-second of container runtime
  • Concurrency: Up to 20 concurrent containers per sandbox worker
  • TTL: Containers expire after 5-60 minutes (configurable per session)

Short-lived sandbox sessions (5-15 minute TTL) keep container costs low. Long-running interactive sessions (30-60 minute TTL) cost more but are less frequent.

  1. Use prompt caching aggressively. Batch sequential AI operations within the same program to maximize cache hits on shared context.
  2. Set appropriate sandbox TTLs. Default to 15 minutes for automated execution. Reserve 30-60 minute TTLs for interactive debugging sessions.
  3. Monitor AI usage records. The aiUsageRecords table tracks every API call with model, token count, and cost. Use the Agent Activity dashboard to identify expensive patterns.
  4. Leverage free tiers for evaluation. Convex Free, Clerk Free (10K MAU), Vercel Hobby, and Cloudflare Workers Free cover evaluation workloads.
  5. Right-size model tiers. Opus 4.6 is reserved for document analysis where extraction quality matters. All structured analysis routes use the more cost-effective Sonnet tiers.