50k tokens, one key, under $1: a fair price check

01 / Compute the 50,000-token session cost before you trust any price page

Most price comparisons fail before they begin because they skip arithmetic. So start with one concrete session and price it twice. Take a coding session of 50,000 real tokens: 32,000 input tokens and 18,000 output tokens. That is a realistic hour of asking a model to read a file, explain it, and rewrite a function.

The formula is short and it is the only one you need for a first pass. Cost equals input tokens divided by one million, multiplied by the input price per million, plus output tokens divided by one million, multiplied by the output price per million.

Worked example with round numbers. Suppose the input price is 1.00 per million tokens and the output price is 5.00 per million tokens. Your input line is 32,000 divided by 1,000,000, which is 0.032 of a million, times 1.00, which is 0.032. Your output line is 18,000 divided by 1,000,000, which is 0.018 of a million, times 5.00, which is 0.09. Add them and one session costs 0.122, or a bit over 12 cents. Under one dollar for that session is not a slogan, it is arithmetic.

Now project it. Ten such sessions a day for 30 days is 300 sessions. At 0.122 each, the month is 36.60. This number is the anchor for every comparison below. If a vendor quote cannot be converted into the same 50,000-token shape, it is not comparable yet.

Two honesty notes. First, use the actual listed prices for the model you pick, because model prices differ widely. Second, this is a planning calculation, not a benchmark and not a discount claim.

02 / Normalize the three price units you will see on every page

Vendor pages quote prices in at least three different units, and mixing them is how budgets get wrecked. Convert everything to one shape: currency per one million tokens.

• ■ Per 1,000 tokens: multiply the number by 1,000. So 0.002 per 1,000 tokens becomes 2.00 per million tokens.
• ■ Per 1 million tokens: use as-is. This is the easiest unit for comparisons.
• ■ Per request or per call: divide the fee by your average tokens per call to get an effective token price, then scale.
• ■ Subscription or monthly plan: divide the monthly fee by the token volume you would realistically consume.

That last line deserves numbers. If a plan costs 20.00 per month and your session costs 0.122, the plan must deliver at least 163 sessions per month to beat pay-as-you-go, which is about 5.5 sessions every day. On a tiny budget, if you only build on weekends, pay-as-you-go wins and the unused subscription is dead money.

Also watch the small print around money before tokens: minimum top-ups, credits that expire, and balances that are non-refundable. A 10.00 minimum top-up is not a price, but it does affect how quickly you can test a service.

03 / Cache reads and cache writes are the lines that double a bill

Long-context work looks cheap on a naive calculation and expensive on a real one. Imagine a 200,000-token codebase prompt that you resend across 20 turns in one session. Counted naively as fresh input every time, that is 4,000,000 input tokens before the model writes a single line of output.

This is where fair comparison matters most. Some providers bill cached prompt tokens at a lower read rate, and bill the first write of that cache at a higher rate. Others round to the nearest block, so a prompt of 21,000 tokens may be charged as 24,000. Those block boundaries can quietly add up over a month.

To compare fairly, tally three quantities, not one: cache write tokens, cache read tokens, and uncached input tokens, each multiplied by its own rate. Then add output tokens at the output rate. Only now can you put two services side by side.

Check how the relay handles this. CheapRelay bills per real token, including cache reads and writes, with no monthly fee and no subscription, so the receipt reflects what the model actually consumed rather than a rounded block. Whatever service you use, demand the same transparency.

04 / One key and compatible routes remove their own hidden costs

On a tiny budget, engineering time is money even when no invoice arrives. If you must sign up for five vendor dashboards, manage five keys, and maintain five client configs, the cheap option can become the expensive option.

A relay collapses that. CheapRelay gives you one key to reach Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Qwen, Grok and more. The API base is https://api.cheapapirelay.com, with OpenAI-compatible routes under /v1 and Gemini native routes under /v1beta. That means it works with Claude Code, Codex, Gemini CLI and any OpenAI SDK, so your existing code keeps running while you switch models.

The practical win is testability. With one key you can run the same 50,000-token prompt shape against two models and compare real receipts instead of list prices. If a cheaper model finishes the task in fewer sessions, the effective cost is lower even when its per-million rate is higher. Count sessions completed per dollar, not just tokens per dollar.

Billing is pay-as-you-go against real token use, so a quiet week costs nothing. There is also a referral reward worth 20 percent of what invited users pay, which helps when a small team shares one budget.

05 / Keep a receipt ledger that turns list prices into one honest line

Finish with a habit, not a tool. For every service you try, write one row per week in plain text: model name, input tokens, cache write tokens, cache read tokens, output tokens, total charged, and tasks completed. Six numbers and one verdict.

• ■ Convert every quote to currency per one million tokens before comparing.
• ■ Split input, cache write, cache read, and output lines instead of trusting one blended number.
• ■ Project the monthly total from your real session count, using 0.122 for a 50,000-token session as a sanity anchor.
• ■ Ignore uptime claims, benchmarks and discount banners that are not on your own receipt.
• ■ Prefer pay-as-you-go when your usage is irregular, and re-check the math only when a subscription volume is genuinely high.

Create a CheapRelay account at cheapapirelay.com and start with the smallest top-up that lets you run ten 50,000-token sessions. That is roughly 1.22 at the example rates above, and it buys exactly the evidence you need.

Compare that receipt against a list price for the same token shape, including cache lines. If the relay line is lower for the work you actually do, you have a fair comparison. If it is not, keep your money and keep the ledger. Either way, the next decision is made with numbers instead of headlines.