How to Cut AI Costs? With the Right Model, the Same Question Can Cost 100x Less

Once a company adopts AI, the biggest variable in its AI costs is often not usage. It is the model it chose.

Take a single customer service question. Assume about 10,000 input tokens (the system prompt, knowledge base snippets and the customer's question) and 500 output tokens. At the rates of MaiAgent, an enterprise AI platform, answering it with GPT-6 Luna costs about 25 Credits. With GPT-6 Astra it costs about 2,475 Credits, a 100x difference.

For a company that answers tens of thousands of questions a month, that is a budget gap of two orders of magnitude. But more expensive does not mean better. Of the 45 text models available on MaiAgent that have a public capability score, 39 have another option that is cheaper and at least as capable.

You can also start with the video version, under 3 minutes long.

1. How are AI costs calculated? LLM pricing comes down to input and output

1. How LLM billing works

Large language models are billed by the token, with input and output priced separately. The Credits deducted for each model call are calculated as follows.

Credits = input tokens ÷ 1 million × input rate + output tokens ÷ 1 million × output rate

Take GPT-6 Luna. Its input rate is 1,980 Credits per million tokens and its output rate is 9,900 Credits. A customer service question with 10,000 input tokens and 500 output tokens comes to 19.8 plus 4.95, or about 25 Credits. The input and output rates for every model are published in Credits Billing System.

2. Three AI costs that are often underestimated

The formula is simple, but three costs are often underestimated.

Still, the biggest factor in AI costs is which model you choose.

2. Comparing LLMs: a more expensive model is not always better

1. What is the efficient frontier?

The efficient frontier is made up of the models that are the most capable among all models costing the same or less. Any model that is not on the efficient frontier has a cheaper alternative that is at least as capable.

The chart below plots each model as a dot. The horizontal axis is the Credits cost of one answer, and the vertical axis is the Epoch AI Capabilities Index (ECI). The data is the same as in the scatter chart in the MaiAgent admin console.

Of the 45 models, only 6 are on the efficient frontier. A few examples show that more expensive does not mean better.

• GPT-6.1 Terra has a capability score of 168.1, just 0.3 points below Claude 5.5 Opus (168.4), at a quarter of the cost. It also costs half as much as Claude 5.5 Sonnet and GPT-6.1 Sol, and scores higher than both.

• Claude 5.1 Fable costs 2.5 times as much as Claude 5.5 Opus, yet scores 2 points lower.

• Claude 5.5 Haiku scores 160.4, already higher than Claude 4.8 Opus (159.6), the flagship model in May 2026, at 1/45 of its cost.

Model prices and capabilities change quickly, much faster than most companies update their settings.

2. The most cost-effective models today

Pick models by task difficulty, moving up a tier as tasks get harder, and in each tier choose a model on the efficient frontier. The table below uses November 2026 rates and assumes 10,000 input tokens and 500 output tokens per customer service question.

With the Professional plan's 5 million Credits a month, using GPT-6 Luna for everything covers about 200,000 customer service questions, while using GPT-6 Astra for everything covers only about 2,000. The Enterprise plan's 10 million Credits a month doubles both numbers.

3. How can you lower AI costs? Start with tiers, then look for savings

1. Tier models by task difficulty

The most direct way to lower AI costs is tiering. Send high-volume questions with clear rules to light models, and use mid-tier or flagship models only for agents that need multi-step reasoning or work across systems. With the same Credits budget, a light model can answer dozens to 100 times as many questions as a flagship model.

2. Where you can still save after choosing a model

Once you have chosen a model, there is still room to save on that same model, for example by shortening the context resent on each call, making good use of prompt caching and adjusting the thinking effort.

To turn these savings into a return on investment, or to see how falling inference costs affect procurement strategy, read these two articles next.

Further reading: "How to Calculate AI ROI? A Framework Enterprises Can Actually Use" and "DeepSeek: Should Enterprises Rethink AI Procurement?"

4. AI cost calculator: run the numbers on your own usage

The table below lists the Credits each model uses per request in three common scenarios. Divide your monthly Credits allowance by the cost per request to see how many requests it covers in a month. For example, the Professional plan includes 5 million Credits a month, enough for about 200,000 customer service questions answered with GPT-6 Luna.

5. How fast are LLM prices falling? The same capability now costs hundreds of times less

1. Same capability, steadily falling prices

The trend is clearer if you hold the capability score fixed and look at what the cheapest model reaching that score cost at each point in time. The chart below converts each model's official launch price into Credits per question.

The newer the capability level, the faster its cost falls. In April 2026, only GPT-5.5 reached the GPT-5.5 level. Six months later, Claude 5.5 Haiku reached it too.

External research shows the same trend. Epoch AI measured six benchmarks and found that the price of reaching a fixed score falls by 9x to 900x per year. a16z calls this LLMflation and estimates that the same capability gets about 10x cheaper each year.

2. Big price drops mostly come from new light models

Big price drops usually do not come from old models getting cheaper. They come from new light models. In July 2024, GPT-4o mini cut the cost of GPT-4-level capability by 21x in one step. In October 2026, Claude 5.5 Haiku cut the cost of the GPT-5.5 level by 20x in one step. So the model you choose today will quite likely have a cheaper equivalent within six months.

3. Open-weight models and on-premises deployment

Open-weight models have caught up too. GPT-OSS 20B is an open-weight model that can run on a single GPU. Its capability score of 137.8 is already above GPT-4 from 2023, which is also why on-premises deployment is becoming more practical.

6. How do you validate before switching models? Retest with your own questions

1. A higher capability score does not mean better performance in your use case

You should switch models regularly, but validate before each switch. A higher capability score does not mean better results on your knowledge base or your customers' questions, and response speed, format and tone may change as well. We have seen cases where a model was switched without testing first and response times got slower. We have also seen a customer rerun a set of nearly 300 test questions and decide not to switch for now.

Further reading: "AI Models Change Weekly. Can Your Company Keep Up?"

2. MaiAgent Serve's standard process for switching models

The MaiAgent Serve operations service turns model switching into a standard process, and every step is recorded.

1. New model released

2. Rerun with your test questions

3. Old-vs-new comparison report delivered

4. Switch only with your approval

5. Ongoing monitoring after the switch

For on-premises customers, model switching is handled by Model Care.

Conclusion

Keeping AI costs in check comes down to giving each task a model that is good enough and good value. Tier by task difficulty first, pick a model on the efficient frontier for each tier, then confirm real-world performance with your own test questions. Model prices and capabilities will keep changing, and regular reassessment is what keeps the same budget delivering the most value.