What an AI Task Costs: How to Estimate It and Choose a Model

Nobody sets out to buy a million tokens. You set out to get a task done: audit a CRM, draft follow-ups for forty leads, clean up a set of records. What it costs depends on how much work the AI needed to finish, and that is easy to misjudge by looking at per-token prices alone. This article explains what drives the cost of an AI task, shows a way to estimate it, and offers a practical way to choose between models. The framework draws on how Anthropic describes task cost in its own published material, and the examples and advice are ours. The method applies to any AI vendor.

No prices here, on purpose. Per-token prices and model line-ups change often, so quoting them would date this article quickly and could mislead you. The worked example below uses made-up ratios to show the arithmetic. Before you budget, take the current price list from your vendor, and use the method here to turn it into a cost per finished task.

A task is a loop, not a single request

When an AI works through a task using tools, such as reading CRM records through Zoho MCP, it goes round a loop. It reads the conversation so far, calls a tool, reads the result, and goes round again until the job is done. Each trip round is one request, or turn. Four things decide what the loop costs.

Driver	Why it matters
Number of turns	Every turn resends the whole conversation so far. A task that needs eight turns processes far more text than one that needs two, even if each turn looks small.
Fresh versus cached input	Most of what a turn resends is text the model already saw. Many vendors bill that repeated text at a small fraction of the normal input price, so it is cheap, but only if the vendor supports it and your set-up allows it.
Output, including reasoning	Output is usually the most expensive kind of token, typically several times the price of input. Where a model reasons before answering, that reasoning is often billed as output too, even if you only see a summary.
The model you choose	Each model has its own prices, and a more capable model may need fewer turns, so a higher price per token does not always mean a higher price per task.

A worked example, in units rather than pounds

To keep prices out of it, count cost in units. Suppose, purely for illustration, that 1,000 fresh input tokens cost 1 unit, 1,000 output tokens cost 5 units, and 1,000 cached tokens cost 0.05 units. These ratios are invented to show the effect, and your vendor's real ones will differ.

The task: review 20 CRM records and draft a follow-up note for each. The instructions and records come to about 12,000 tokens.

Approach	What happens	Illustrative cost
One turn	12,000 fresh input tokens (12 units) and 600 output tokens (3 units).	About 15 units
Six turns, using tools	The same 12,000 fresh at the start (12 units), then five more turns that each resend the growing conversation from cache (about 75,000 cached tokens, 3.75 units), add about 1,000 fresh tokens of tool results each (5 units), and produce 600 output tokens each (6 turns in total, 18 units).	About 39 units

Three lessons carry over to real prices. The multi-turn version costs roughly two and a half times as much, because the loop repeated the work. Output made up nearly half of the total, even though there was far less of it than input. And the cache made the repeated reading cheap, so it was the turns and the output that cost the money, not the resending.

Where the cost shows up depends on how you connect

Route	How you pay and how to watch it
An AI assistant on a subscription plan	Usually a per-user fee with usage limits set by the vendor. The cost is predictable, but heavy automated use can hit the limits.
Your own API key, or a cloud platform	Billed by usage, so the method above applies directly. Watch the vendor's own usage dashboard.
Zoho running models for you	Through Zoho Key Service, you spend a pool of credits, and Zoho CRM's Usage and Limits tabs show where tokens and credits go and let you cap spend.

Cutting cost without cutting results

Narrow the task. Fewer records and fewer tools means a smaller conversation and fewer wrong turns. See Connecting Claude to Zoho: A Safe Setup Checklist and Prompting Guide.
Ask for what you need and no more. A request for a table of five fields costs less output than a request to explain everything it found.
Keep instructions stable. Where your set-up supports caching, a fixed set of instructions that does not change from one request to the next is the part most likely to be cached.
Batch similar jobs instead of starting a new conversation for each record, where the task allows it.
Set limits before you need them. Use your vendor's spend limits, or Zoho CRM's Limits tab, so one heavy user cannot run up the bill unnoticed.

The false economy to avoid: a cheaper model, lower effort, or less context can save money on every turn and still cost more overall, if it means the task takes more turns, or fails and has to be run again. Judge cost by the price of a finished task, not the price of a turn.

Choosing between models

Most AI vendors offer a ladder of models, from small and fast to large and capable. Think in tiers rather than names, because the names change. This is our own rule of thumb, not a guarantee.

Type of task	Tier to try first
Simple, high-volume work with clear rules: classifying, extracting fields, tidying text	The smallest, cheapest tier.
Everyday drafting and summarising, with a person reviewing the result	A middle tier.
Multi-step work with tools, audits, analysis, or anything that is hard to check	The most capable tier, where fewer wrong turns often pay for the higher price.
Anything that writes to live data	A capable tier, plus a plan-first prompt and human review, whatever the cost.

Test with your own work

Pick around ten typical tasks, covering the easy, the average, and the awkward.
Run the same tasks on two tiers, with the same instructions.
Score each result as finished correctly, needed correction, or failed.
Divide the total spend by the number of correctly finished tasks. That is your cost per finished task, and it is the number to compare.

Zoho CRM lets you assign different models to different teams, which suits this approach: see Manage Your LLMs with Model-Level Configuration. For our overall recommendation on which AI to use with Zoho, read Using AI with Zoho: Why We Recommend Claude, and When Something Else Fits Better.

Need help budgeting an AI project? As a Zoho Partner serving the UK and Ireland, we can estimate the cost per finished task for your use case, set up limits, and test model tiers against your own data. Email zoho@1cloudconsultants.com or raise a support request with 1 Cloud Consultants.