$3.85 API Receipt, $20 Chat Subscription

01 / A $20 subscription needs a usage receipt

If you pay $20 a month for web chat but use it only occasionally, an API can cost less. It can also cost more. The difference is not the number of days you open the app; it is the volume of billable input, output and cache activity your workflow generates.

Here is a worked migration budget for 100 tasks per month. The $20 subscription is an assumed comparison price, not a claim about every chat plan. All token rates below are illustrative arithmetic inputs, not advertised CheapRelay prices. Replace them with the current rates for your chosen model before spending money.

02 / The sample receipt totals $3.85 for 100 tasks

Assume each task averages 20,000 uncached input tokens, 4,000 output tokens, 5,000 cache-read tokens and 1,000 cache-write tokens. Those are separate billing buckets in this example: do not count the same input tokens twice.

• ■ Uncached input: 2,000,000 tokens at an assumed $1 per million = $2.00.
• ■ Output: 400,000 tokens at an assumed $4 per million = $1.60.
• ■ Cache reads: 500,000 tokens at an assumed $0.20 per million = $0.10.
• ■ Cache writes: 100,000 tokens at an assumed $1.50 per million = $0.15.
• ■ Total: 3,000,000 billable tokens across all buckets = $3.85.
• ■ Average API charge per task: $0.0385, or 3.85 cents.

Against the assumed $20 subscription, that leaves $16.15 unspent, or 80.75% of the subscription price. This is a scenario calculation, not a promised saving. It excludes any separately charged tools, client software, hosting, taxes or payment costs that might apply.

03 / The same workload crosses $20 at 520 tasks

At an unchanged 3.85 cents per task, 500 tasks cost $19.25. At 520 tasks, the bill reaches $20.02. That is the useful comparison: this particular workload becomes more expensive than the assumed subscription at roughly 520 tasks per month.

Real tasks rarely stay identical. A coding agent may make several requests before finishing one job. A long conversation may resend earlier messages. Retries and larger outputs can increase billed usage. A cache hit may lower the applicable input charge, but cache reads and writes are not automatically free.

Do not translate “100 chats” into “100 API requests.” Measure complete jobs, including follow-up calls, so the receipt reflects what you actually wanted accomplished.

04 / Move ten real jobs before canceling web chat

CheapRelay is a pay-as-you-go AI API relay with no monthly fee or subscription. It bills per real token, including cache reads and writes. One key provides access to multiple model families, which can simplify testing without maintaining separate credentials for each provider.

Sign up for CheapRelay and start with ten representative jobs rather than moving your entire workflow immediately. Choose a model, check its current rates, and record input, output and cache usage for every completed job.

For an OpenAI-compatible client, use https://api.cheapapirelay.com/v1 as the API base. Gemini-native routes are under https://api.cheapapirelay.com/v1beta. CheapRelay works with Claude Code, Codex, Gemini CLI and OpenAI SDK clients. Confirm the client’s expected endpoint format before sending requests.

05 / Keep software and migration time on the receipt

An API supplies model access, not necessarily the complete web-chat experience. You may need a client for conversation history, file handling or a convenient interface. If that client costs $5 monthly, the example total becomes $8.85, leaving $11.15 against the $20 baseline.

Your time belongs on the receipt too. One hour of setup valued at $25 exceeds the first month’s $16.15 usage saving. That does not make migration pointless, but it changes the payback period.

Cancel the subscription only after the trial confirms acceptable results, complete workflow coverage and a lower total bill. Keep whichever arrangement makes your actual receipt smaller, not whichever pricing headline looks cheapest.