$8.46 a Month to Retire a Hobby App’s Free Tiers

01 / The $8.46 receipt for a small AI-powered app

Moving off free tiers should buy fewer constraints, not a subscription collection. For a hobby app that summarizes saved articles, I would start with a monthly planning budget of $8.46: $5.46 for infrastructure and $3.00 for AI requests.

This is a worked estimate, not an actual customer bill or a quote from CheapRelay. Every unit rate below is an illustrative budgeting assumption. Replace those rates with the current prices of your chosen services before depositing money or deploying anything.

The workload is 2,000 completed summaries per month, roughly 67 per day. The app has one small worker, a database on its attached persistent storage, a basic backup job and no paid analytics service.

02 / Infrastructure receipt: $5.46 before AI

• ■ Compute: 500 running hours × $0.008 per hour = $4.00.
• ■ Persistent storage: 10 GB-month × $0.10 = $1.00.
• ■ Backup storage: 4 GB-month × $0.10 = $0.40.
• ■ Outbound traffic: 3 GB × $0.02 per GB = $0.06.
• ■ Infrastructure subtotal: $5.46.

The 500-hour compute allowance assumes a worker that stops during quiet periods. An always-on deployment changes that line: 720 hours in a 30-day month would cost $5.76 at the same assumed rate, raising the full monthly total to $10.22.

Check whether stopped instances still incur storage charges, whether a public IP costs extra and whether backups have minimum charges. This receipt excludes domain registration, taxes and payment-processing charges. An existing domain is not free merely because its renewal happens next year.

03 / AI receipt: $3.00 across four token categories

For each summary, assume an average of 1,000 uncached input tokens, 300 output tokens, 3,000 cache-read tokens and 160 cache-write tokens. These are separate accounting buckets, not four names for the same input. Confirm your provider’s usage definitions before applying this calculation.

• ■ Uncached input: 2 million tokens × $0.50 per million = $1.00.
• ■ Output: 600,000 tokens × $2.00 per million = $1.20.
• ■ Cache reads: 6 million tokens × $0.10 per million = $0.60.
• ■ Cache writes: 320,000 tokens × $0.625 per million = $0.20.
• ■ AI subtotal: $3.00, or $0.0015 per completed summary.

Those rates are arithmetic examples, not advertised model prices. Cache usage is also an assumption, not a promised hit rate. If a cache expires or requests cannot reuse a shared prefix, recalculate using the actual uncached input and cache-write totals.

CheapRelay fits this approach because it bills per real token, including cache reads and writes, without a monthly fee or subscription. One key provides access to multiple model families, and its OpenAI-compatible routes work with an OpenAI SDK. That can simplify a small app’s integration without turning access itself into another fixed monthly expense.

04 / The migration starts with 100 paid summaries

Move one background task first, rather than replacing the entire stack. Keep the free deployment available while you test the paid path. Record usage by request, including failures and retries; a successful summary counter alone cannot explain an API bill.

At the assumed usage, 100 summaries cost $0.15 in AI charges. Compare that estimate with reported usage, then replace every assumption with observed averages. Sign up for CheapRelay when you are ready to test that small paid batch, and check current model pricing before choosing a model.

05 / A $12 ceiling leaves $3.54 for mistakes

Against the $8.46 estimate, a $12 monthly operating target leaves $3.54 of headroom. Treat that as a warning margin, not permission for unlimited retries. Set alerts below the ceiling, cap output length and limit each job to one automatic retry.

At unchanged average token usage, doubling volume to 4,000 summaries makes the AI line $6.00 and the total $11.46, assuming infrastructure usage stays flat. In practice, storage, traffic or compute may grow too.

Keep a daily usage ledger and pause optional jobs when spending outruns the plan. Provider alerts are useful, but they are not necessarily hard spending caps. The point of leaving free tiers is a more predictable project, not a larger surprise.