AIStall
EditionBy The AI Stall editorsNovember 1, 2026

How to Size an AI API Account Before You Pay

How to Size an AI API Account Before You Pay

Choosing an AI platform by its per-token price answers one question out of six. The other five are also published, and they are also numbers.

What this note covers, in order.
What this note covers, in order.

This note collects the checks worth doing before an API key gets committed to a product. Every figure was read from vendor documentation on October 27, 2026.

ChatGPT Business's public homepage, captured automatically on October 10, 2026. Not a logged-in account; the page may have changed since.
ChatGPT Business's public homepage, captured automatically on October 10, 2026. Not a logged-in account; the page may have changed since.

It draws on four earlier notes in this brief and adds the account-level terms that sit before any of them: data use, tier qualification, model availability and tool billing.

Read what the free tier is paying you with

The Gemini API pricing page sets its free and paid tiers side by side, and the last line of each column is the one to read first.

The free column reads: content used to improve our products. The paid column reads: content not used to improve our products. Both carry a footnote marker.

The free column also says limited access to certain models. The paid column lists higher rate limits, access to context caching, the Batch API at a 50 percent cost reduction, and access to the most advanced models.

So the free tier is not a smaller version of the paid tier. It is a different agreement with a different data clause, and the clause is printed on the price page rather than buried.

What each vendor says it keeps, and for how long, is collected separately in the retention note.

The step up is a purchase on one platform and a wait on another

All three platforms gate throughput behind a tier. How a new account climbs differs enough to change a launch plan.

The waiting periods are the part a schedule has to respect. Google's Tier 3 requires thirty days to have passed since the first successful payment, whatever the amount since.

Anthropic adds a caution in the other direction. New organisations may start in an Evaluation tier, with limits below the standard published ones until account history exists.

Google adds its own. Meeting the stated qualification criteria is generally sufficient for approval, but in rare cases an upgrade request may be denied after review.

The ceilings those tiers carry are in the monthly cap note and the throughput they buy is in the rate limit note.

A new project cannot always pick the cheap old model

Model availability is not the same as model existence, and the Gemini models page says so in a repeated note.

Access to the 2.5 models is being limited to users who have actively used them in the past. The note says they are not deprecated and will continue to be served until further notice.

For any new projects, the page directs readers to Gemini 3.5 Flash-Lite or Gemini 3.8 Flash instead. The stated reason is maintaining capacity for both legacy workflows and new applications.

That is a real constraint on a cost model built from an older price row. A team that budgeted against Gemini 2.5 Flash cannot necessarily call it from a new project.

The same page notes that Gemini 4 Argon is not available on the free tier at all. Evaluating the flagship model requires a billing account first.

Retirement dates for models that are callable today are collected in the model shutdown calendar.

One catalogue has four models and one has over a hundred

The two documentation sites present radically different choices, and the difference is countable.

The OpenAI models page lists four featured models above a full catalogue. Counting the catalogue entries on the markdown version of that page gives 103 separate model listings.

The Anthropic models overview compares four current models in one table: Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. Specialised and legacy models sit in their own sections.

Neither approach is wrong. A long catalogue keeps older snapshots callable, which is what a production system pinned to a version needs.

It does change the first hour of work. On one platform the decision is which of four rows to read. On the other it is which of a hundred entries is still the right one.

Both vendors give a shortcut. OpenAI names a default flagship on the page itself, and Anthropic's page says to start with Claude Opus 5.5 for most workloads.

Grounding and tools bill separately

A token budget is not a bill. Tool calls are metered on their own lines and the rates are not per token.

The Gemini pricing page lists grounding with Google Search at 5,000 free search requests per month, shared across all Gemini 3 and newer models, then $14 per 1,000 requests.

Grounding with Google Maps carries the same structure: 5,000 free requests per month shared across the same model generations, then $14 per 1,000 search queries.

The shared allowance is the detail to note. Adding a second grounded model to a project does not add a second 5,000 requests.

Anthropic's pricing page carries its own separate sections for tool use, specific tools and Managed Agents, including session runtime as a billable quantity rather than a token count.

The order to check things in

Bottom line

The per-token rate is the most visible number on an AI platform and rarely the one that decides a bill.

A monthly ceiling, a tier qualification, a long-prompt threshold, a throughput limit and a tool-call rate all sit around it, and all five are published.

The work is reading them in one sitting before the first key is issued, rather than one at a time after each surprise.

The four notes behind this one: the monthly cap, the long-context price threshold, the million-token window and the rate limit tables.

Teams buying seats rather than keys should start with the ChatGPT Business review and the shortlist of AI assistant plans.

What this brief prints and what it cuts is in issue zero.

Sources