Tokens, effort, and cost

AI usage is billed in tokens. Knowing how tokens add up lets you budget AI per sheet instead of finding out from the invoice.

Tokens​

A model reads and writes in tokens: chunks of text that average about three-quarters of a word in English. A short phrase like expansion joint cover is a handful of tokens. A dense spec page is about a thousand. A general-notes sheet, parsed to text, is two to three thousand.

Everything the model reads is input tokens. Everything it writes is output tokens. Output tokens cost several times more than input tokens.

In a conversation with several prompts, earlier context is re-sent to the model with every new prompt. Providers discount context they have already seen, called cached tokens, so a long session costs less per prompt than the raw count suggests.

Open figure

Effort​

Effort is how long the model reasons before acting. Four levels:

Effort	Effect
Low	Faster. Reasons just enough to answer or pick a tool.
Medium	Reasons longer before acting. Prefers a complete answer over speed.
High	Thorough reasoning. Larger tool budget.
Extra High	Longest reasoning and largest tool budget. Costs the most.

Higher effort means more reasoning tokens and more tool calls. Higher effort costs more and takes longer. In return the model reads more of the set before deciding and checks itself more. Auto keeps effort at Medium. Drop to Low for a quick lookup; raise it for a milestone review. In Nomic, effort can change mid-session; the model cannot.

For comparison with other tools: a deep-research mode is roughly Medium, and a long-running background agent is roughly High.

Open figure

How Nomic bills​

One AI Usage pool, in dollars. Every AI operation deducts from it at the provider's rate:

• Agent and workflow runs: per million tokens at the model's API price. Input, output, cache read, and cache write are metered separately, plus a per-million Nomic Token Fee
• File parsing: per page, by Nomic's Parse model, once for each version of a file

Plans differ in where usage is managed:

• Free: limited agent requests instead of a dollar pool. Upgrade under Settings → Billing.
• Individual: a personal pool, under Settings → Billing.
• Business and Enterprise: pooled across the organization, under Admin → Usage.

Rates are the same on every plan. See Models & Pricing.

What things cost​

Approximate, on the default model (Sonnet 5.5), from Plans:

Task	Estimated cost
One cited question against a project's files	~$0.10 – 0.15
Typical Nomic agent session (Low effort, 5–10 messages)	~$0.10 – 0.25
Drawing review, per sheet	~$0.25 – 1.00
Medium-effort research across a large project	~$0.50 – 1.50
Submittal review against a project spec (multi-file, structured)	~$0.75 – 2.00
Code compliance check against a screenshot from a sheet	~$0.15 – 0.40

Cost tracks how much the model reads, how many tool calls it makes, and how long the answer is. A drawing review reads every sheet and writes a finding per issue. A lookup, such as finding which spec section covers roof membrane warranties, reads one page and writes a sentence.

Before a workflow runs, Nomic shows a spend estimate: a dollar range from the size of the inputs and the cost of past runs. The estimate sharpens with use. Because it is a forecast, the actual deduction can differ, so use it as your chance to check the scope before the run starts.

Open figure

Keeping spend down​

From Models & Pricing:

1. Default to Sonnet 5.5 and Low. Raise effort for milestone reviews.
2. Keep projects scoped. A smaller project, folder, or workflow context means less input per prompt.
3. Reuse files already in the project. Parsing is charged once per page of each file version, so uploading the same file twice pays twice.
4. New task, new session. A forty-message drawing review carried into an unrelated RFI question is re-read and billed on every new prompt.

Next: what the model should be reading in the first place.

Next: Context