← jwatte.com

Model-Tier AI Cost Calculator

Put your own task in and see what it costs per month on every Claude, OpenAI, and Gemini tier, then see how much you save by routing it: an expensive model plans, a cheaper one does the writing, and the cheapest one verifies. Pick a preset to start, share a link to any estimate, and edit any price. Everything runs in your tab. Background: why the priciest model should not be your task runner.

Your workload

Start with a preset, or enter your own

Pick a task shape to fill in typical numbers, then adjust. A preset just sets the three fields below.

Input is everything you send the model: your instructions plus any context or documents. Output is what it writes back. A token is roughly three quarters of a word, so 1,000 tokens is about 750 words. The word counts under each field update as you type.
Cached tokens bill at roughly 10% of the input price. The stable stuff (a system prompt, a rulebook, a document you reuse) is what caches.
Route it

Pick the model for each job

The routed plan gives each step to the cheapest model that can do it: the orchestrator reads the whole task and writes the plan, the worker writes the bulk of the output, and the verifier checks it. Mix providers freely.

The comparison

One model does the whole task

The same workload run entirely on each model. This is what most people do: pick one model and point it at the job.

Model$ / run$ / monthRelative

Route it across tiers

RoleModelIn tokOut tok$ / month
Routing splits the expensive output across tiers and leans on the cache for the shared context, so the priciest model only ever writes the plan. On an input-heavy task with little caching, routing saves less; turn the cache slider down and watch it. The interest is in the output half, and caching is what tames the input half.
Advanced: prices and the routing split

Prices are US dollars per million tokens (standard text rate, not cached or batch). Defaults are approximate mid-2026 list prices; change any to match today.

The worker writes whatever the plan share leaves. The verifier reads the produced output fresh (no cache) and returns a short verdict.

How this maps to the loop

The routed row is the pattern the companion post lays out: the priciest tier plans and judges, cheaper tiers do the typing, and the cheapest tier gates the output. The calculator only prices it. The full setup, with the memory and skill files you can copy, is in the write-up. The other side of the cost question, running models on your own hardware instead of renting them, is in where APUs beat GPUs for on-prem AI.

Prices change often and vary by tier. Defaults are standard text list rates checked against each provider on July 17, 2026. Claude Sonnet 5 bills at an introductory $2 in / $10 out through August 31, 2026, shown here at its standard $3 / $15. Some models charge more above a long-context threshold, such as Gemini Pro above 200k tokens, and most offer cached-input and batch discounts, which are simplified here. Confirm current numbers on the provider pages before you budget: Anthropic, OpenAI, Google Gemini. These estimates are illustrative and assume a simple caching and routing model.