You overpay for AI almost tenfold. Here is how to fix it

When a founder tells me AI is eating their margin, I ask to see one thing: which model answers a simple question like where is my order. The answer is almost always the same. The most expensive one. The same one that writes the complex analysis.

That is delivering pizza in an armored limousine. It gets there, but the economics never work. The good news is that most of the overspend goes away with four settings, and none of them require rewriting your product.

Every number below comes from Anthropic's official pricing as of September 2026. Other providers follow the same logic, only the rates differ.

Start with the price list#

• Claude Opus 5: $5 per million input tokens and $25 per million output tokens.
• Claude Sonnet 5: $2 and $10. Anthropic also dropped the increase to $3 and $15 that was scheduled for September 1, so the introductory price is now the standard one.
• Claude Haiku 4.5: $1 and $5.

Two things jump out. The gap between the top and the bottom model is fivefold. And output tokens cost five times more than input tokens across the board. Every saving below grows out of those two facts.

Let us do the math on a real case#

Take a support bot handling 100,000 conversations a month. Each request carries 8,000 tokens that never change (instructions, knowledge base, tool descriptions), 500 tokens of the customer's question and 400 tokens of answer.

The default setup: everything goes to Opus 5. Input is 850 million tokens for $4,250, output is 40 million tokens for $1,000. Total: $5,250 a month.

Now apply the settings one at a time.

Setting 1. Send the easy stuff to a cheap model#

In support, usually four questions out of five are routine: order status, returns, address changes. Haiku handles them fine. Hard cases, complaints and anything the model is unsure about go to Opus. The cheap model itself or a simple rule set makes the call, and when in doubt the request moves up a level.

The math: 80% of traffic on Haiku costs $840, the remaining 20% on Opus costs $1,050. Total: $1,890. Already almost three times cheaper, and quality on hard questions did not suffer, because a strong model still handles them.

The same principle works across providers: routine work can go to the cheapest suitable model on the market and hard work to the best one. Just make sure the router is your own and not a third-party service. As Anthropic's latest threat report shows, conversations and keys leak through exactly those middlemen.

Setting 2. Stop paying twice for the same thing#

Those 8,000 tokens of instructions and knowledge base are identical in every request, and without caching you pay for them 100,000 times. With caching, the model reads them from memory at 10% of the normal input price. Writing to the cache costs 1.25 times the normal input price for a five-minute cache and twice the price for a one-hour cache. So a five-minute cache pays for itself on the second request.

It is nearly free to build: in Anthropic's API you add a single cache_control field at the top level of the request and the system places the cache breakpoints for you. There is one rule: put everything static at the start of the request and everything that changes at the end.

The math: Haiku now costs $264 and Opus $330. Total: about $600, plus a little for cache writes. Against the original $5,250, that is almost nine times less.

Setting 3. Batch everything that is not urgent#

Overnight ticket tagging, summaries, catalog translation, document classification. If you do not need the result right now, send these jobs through the Batch API. It takes 50% off both input and output, and the discount stacks with caching. For work that can wait, that is one more cut of the bill in half.

Setting 4. Shorter answers#

Since output costs five times more than input, the most expensive token is an unnecessary word in the answer. Ask for a structured format instead of an essay, set a sensible length limit and drop the polite preambles nobody reads. Cutting an answer from 400 to 250 tokens shrinks the output part of the bill by almost 40%.

Three traps where money leaks#

The new tokenizer. Claude models from version 4.7 onward use a tokenizer that, according to Anthropic, produces roughly 30% more tokens for the same text. When you migrate to a newer model, recalculate the budget instead of comparing only the price per million.

The bottomless context window. A one-million-token context is billed at the standard rate on current models, which makes it tempting to stuff all your documentation into every request. Do not. Load only the fragments you need, the rest is money thrown away.

Agent tools. Web search through the API costs $10 per thousand searches on top of tokens. Fetching pages has no separate fee, but everything fetched turns into input tokens. An agent that runs ten searches per question quickly becomes the largest line on your bill.

Where to start tomorrow#

• Track tokens for every feature of your product. The API response already includes input, output and cached token counts.
• Find your three most expensive features and check which model serves them.
• Turn on caching for every request with long instructions.
• Move routine requests to a cheap model and compare quality on a hundred real examples.
• Move everything that can wait into batch processing.

Anthropic's own documentation recommends exactly this: simple tasks on Haiku, most production workloads on Sonnet, the hardest reasoning on Opus, plus caching and batching. The only difference is that most teams scroll past that advice.

AI is not expensive. Sending every question to the smartest person in the company is.