Free tool

Frontier LLM vs task model cost

For one high-volume agent loop: what you pay the frontier API today versus serving a small open-weight task model you trained on your outcomes. No signup required.

Agent calls / day

High-volume loop your frontier LLM currently serves.

Frontier price class

Avg input tokens / call

Avg output tokens / call

Task-model GPU $ / month

0 if you already have spare capacity; else e.g. 200–400 for one mid GPU.

Training recipe (payback)

Labeled rows to train

Train a task model

Assumptions (public)

Frontier
List price presetsInput/output $ per 1M tokens for GPT/Claude-class APIs — illustrative market rates, not a quote from any vendor contract.
Task model
Serving $ you setDefault $0 assumes spare capacity; enter GPU $/mo when you dedicate hardware. Shared inference list rates are on Pricing.
Train
Recipe estimateOne-time LoRA/GRPO estimate from recipe + row count on Shared. Payback = train $ ÷ monthly savings (days). Not a production SLO.

What to do with the number

If savings are material

Start a Pilot on Shared cloud — prove beat-frontier on your eval before BYOC.

Pick a recipe

Browse task model recipes for fraud, decision/risk, and scoped jobs your agent already runs.

See a worked example

Read the dogfood case study — train → eval → swap the weak call.