LLM cost reduction: lower AI costs without losing results

Start with costs by feature

Break down AI spending by feature and check whether it leads to results. ABTO records the model, token usage, cost, and response time for each call.

The same call volume can cost up to 100x more depending on the model

At 100,000 calls a month, the cheapest model (Claude Haiku 5.5) costs $25 and the most expensive (GPT-6 Astra) costs $2,500.

Unless you check that the pricier model earns back the difference in outcomes, that difference is simply cost.

Each request is counted as 1,000 input tokens and 300 output tokens. Input and output are each provider's public price in USD per 1M tokens.

Choose a cheaper model when outcomes hold up

Give some users a cheaper model and check that outcomes such as purchases and conversions hold up before switching. Once it meets your criteria, gradually increase its share and reduce the existing variant to 0%.

If the existing model performs better, you have numbers to judge whether the difference is worth its cost.

Prompt tokens cost money too

Removing unnecessary parts of a prompt can lower cost per request without changing the model. After shortening it, check that your success metrics remain steady.

What teams actually saved

Nonsuri reduced reasoning effort and shortened explanations, cutting model costs by 41% while increasing the payment rate by 18%.

Career SSOT uses different model settings for cover letter drafting and resume editing, cutting model costs by 28% while raising its return visit rate by 12%.

Find the cause of a sudden cost increase

When costs spike, narrow the investigation from a feature to a variant to an individual call.