The Sticker Price Is a Lie, and Your AI Budget Knows It

Your AI vendor's pricing page is telling you a nice, clean number. That number is not your bill. Between promo rates that expire, tokenizers that count your code as more tokens, fixed infrastructure that dwarfs the model spend, and reasoning settings that quietly triple your output, the gap between the sticker and the real cost is wide right now. Let me catch you up on where the money actually goes, and what to do about it before your next margin review.

The number on the pricing page has an expiration date

Anthropic launched Sonnet 5 at $2 per million input tokens and $10 per million output through August 31, 2026. On September 1, those become $3 and $15. That is a clean 50% jump on every token, no auto-extension, one hard date.

The second part is sneakier. Sonnet 5 ships with an updated tokenizer, and the same input maps to roughly 1.0 to 1.35x more tokens depending on content. Plain English prose might be 1.05x. Code, structured data, and non-English text push toward 1.35x. So if your workload is code-heavy, you are looking at the price bump plus a token inflation on top of it. The only way to know your number is to run your own traffic through the token-counting API.

Bring the real date and the real multiplier to your finance conversation. A model that looks cheaper today can cost more than a competitor in September.

The tokens are not the bill

Here is the part that resets the whole conversation. One engineer published the full ledger for Trendvesting, an AI side project running in production since early 2024. His finding: the biggest cost of a small production AI system is not the model tokens. It is the fixed infrastructure that keeps it reliable, the Kubernetes clusters, backups, and observability. Fully loaded, each signal costs about $1.50 to $2.00, and that number is dominated by fixed cost at small scale.

He names three taxes that ate his budget. Paying for bad or free data. Duplicating work because he had no caching or dedup. And wasting money on verbose output, since input and output pricing is asymmetric and long answers are a direct charge. A March move to cheaper models barely dented signal quality once he paired it with disciplined stop-loss execution.

The lesson for your roadmap: when someone models a new AI feature off the pricing page, they are modeling the smallest line item. Ask for the loaded cost per accepted output, not cost per API call.

The setting nobody budgets for

Model choice used to be the whole decision. Now you also pick how hard the model thinks, and that dial spends real money. Reasoning tokens are invisible but billed as output tokens and eat your context window. Give a simple lookup the same thinking budget as a risky production decision, and you pay premium rates to overthink a task that needed a clean answer.

The fix is a four-bucket policy: low for high-volume, easy-to-verify work; medium as the normal default; high for real ambiguity; max for rare, human-gated calls like incident response. Then build an escalation ladder instead of one guess. Start cheap, validate, and only bump the effort when tests fail or confidence drops.

The metric to track is cost per accepted answer, not cost per call. A cheap answer that fails review is expensive. If your team can't tell you why a request used high effort, they can't tell you if it was worth it.

Cheap on paper, expensive in practice

This pattern repeats everywhere. Look at speech-to-text. AssemblyAI Universal-2 lists at $0.15 per audio hour, but a $0.15 route can cost more than a $0.60 route once you add channel billing, paid features, retries, and human correction. A 60-minute stereo call may bill as two hours. Diarization and redaction stack on top. And human correction can beat the API charge by an order of magnitude.

Routing beats picking one winner. One provider for live English calls, another for multilingual meetings, a lower-priority batch route for archives. The advisor pattern in Claude Code makes the same bet a different way, cutting API bills by 85% by using a cheaper model as the workhorse. The warning attached to it is the one to remember: it can introduce silent failures that are harder to debug than just running the expensive model solo.

When your team proposes the cheapest vendor, ask what breaks at 2am and who pays to fix it.

The deep cut

The cost story is moving in your favor faster than the pricing pages suggest, if you measure it right. When OpenAI cut GPT-5.6 Luna prices by 80%, Paweł Huryn ran 105 hidden bugs through it and Anthropic's Fable 5. Luna at max effort fixed 33 bugs for $1.80. Fable fixed 29 for $104. Same class of work, a spread that ran 20x to 90x depending on the session.

That is the practical payoff. The expensive default is often a habit, not a requirement. Before you approve a bigger model tier or a premium plan, run your own benchmark on your own tasks and measure cost per accepted result. Huryn found $20 a month now covers serious work he used to pay $200 for. Meanwhile Azeem Azhar's team modeled AI adoption and found it takes eight years to see which approach actually pays off, and the eventual winner looks like the loser in year five. So the discipline is not chasing the cheapest number. It is knowing your real loaded cost well enough to tell a compounding investment from waste.

Three questions for your team

1. What is our fully loaded cost per accepted output for our top AI feature, including infrastructure and human correction, not just tokens? If we can't answer, that's the first thing to fix.
2. Which of our AI calls run at high or max reasoning effort, and can each one show a test result proving that effort changed the outcome?
3. Do we have a model or vendor benchmark on our own traffic, and when did we last rerun it against the current price cards, including the Sonnet 5 September jump?