Creating an API key usually costs nothing, because providers bill the tokens your requests use. A fair api key cost comparison therefore starts with each provider's official pricing page and ends with cost per task. Check the rates again on the day you act, since model prices change often.
- API key cost comparison is a token price comparison, because providers bill usage rather than key creation.
- Against the published rates checked on September 30, 2026, mid-tier models cost about $2 per million input tokens and $10 to $12 per million output tokens.
- Output tokens cost far more than input tokens, so answer length is often the largest cost you control.
- Caching, batch mode, and routing easy prompts to small models lower the bill before you change providers.
- Compare candidates by cost per task on real prompts, because retries, tool calls, and extra turns can erase a low list price.
API Keys Are Free, but the Tokens They Send Are Not
An API key is a password, not a product. Providers do not charge you to create one. They charge for the tokens each request sends and receives, so an api key cost comparison is really a token price comparison.
Compare Prices on One Page, in One Unit
Open each provider's official pricing page and note the date you checked it. Rates change, and third-party tables often lag. Use one unit for every price: US dollars per million tokens.
| Model | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| gpt-5.6-terra | $2.00 | $12.00 |
| gpt-5.6-luna | $0.20 | $1.20 |
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| Claude Haiku 4.5 | $1.00 | $5.00 |
The table uses published rates from the OpenAI and Anthropic pricing pages checked on September 30, 2026. Prices are examples for this comparison and may already have changed, so confirm them before you promise a budget.
Anthropic pricing and cache rates
Turn Token Prices into Cost per Task
Token prices alone do not tell you which key is cheaper. A cheap model that needs three turns can cost more than a strong model that answers in one turn.
- Pick one real task, such as summarizing a single support ticket.
- Run it ten times on each candidate and log input and output tokens.
- Multiply each token count by the price on the official page.
- Divide the total by ten to get the cost per task.
For example, a task that uses 4,000 input tokens and 800 output tokens costs about $0. 018 on a model priced at $2 and $12 per million tokens. On a $0.
20 and $1. 20 model, the same task falls to about $0. 0018.
These figures are examples, and long-context tiers can cost more.
Which Levers Actually Move the Bill
- Route easy prompts to a small model and keep the large model for hard ones.
- Put stable instructions at the start of the prompt so cached input rates apply.
- Send work that can wait through the batch tier, which costs less per token.
- Cap output tokens, because output costs far more than input on every model.
- Watch cache writes, which can cost more than normal input tokens.
When the Cheaper Key Costs More
- Failed calls and retries still bill, so error loops quietly add cost.
- Tool calls such as web search or code execution add fees beyond tokens.
- A cheap model that needs extra turns can end up costing more.
- Swapping keys means re-testing prompts, which costs engineer time.
Re-check official prices each month and compare cost per task before you switch.
Sources
- Pricing OpenAI · 2026-09-30
- Pricing Anthropic · 2026-09-30
- Claude Opus 5.5 model overview Anthropic · 2026-09-30