Prices for AI models in 2026 are lower than ever — and yet AI projects blow their budgets on a regular basis. The reason: anyone who equates "AI costs" with "token price" is planning around roughly 80 percent of the bill. This article breaks the costs down into their real components, cites documented prices — and works through two concrete examples of where the money actually goes.
The short answer: For most applications, the pure model price (tokens) is the smallest cost block — often just a few euros a month. What eats the budget is integration, maintenance, human review, and data work. Rule of thumb: work out the token price, then reckon the first year's total cost at ten to fifty times that — depending on how deeply the AI reaches into your processes.
Three levels on which AI costs money
Before we look at numbers: AI costs arise on three levels — and only the first appears on the provider's price list.
1. Usage cost (tokens). What you pay the model provider per request. Transparent, billed per million tokens, easy to calculate — and the part everyone talks about.
2. Build cost (one-time). Integration, development, data preparation, setup — the effort until the AI even works usefully in your process.
3. Operating cost (ongoing). Maintenance, human review, hosting, monitoring, error correction, ramp-up. This level keeps running as long as the system runs — and is almost always underestimated.
What tokens really cost
Let's start with the visible part. All prices below are list prices per one million tokens (input / output), as of 2026 — some of them introductory prices. Important up front: output tokens cost a multiple of input tokens. For tasks that generate a lot of text, the output price therefore dominates the bill.
| Model | Price In / Out (per 1M) | Tier |
|---|---|---|
| Google Gemini 3.5 Flash | $1.50 / $9 | Budget tier (volume, routine) |
| Claude Sonnet 5 | $3 / $15 1 | Workhorse (most of the work) |
| Google Gemini 3.1 Pro | $2 / $12 2 | Reasoning & multimodal |
| Claude Opus 4.8 | $5 / $25 | Top tier (the last few percentage points) |
| OpenAI GPT-5.5 | $5 / $30 | All-round flagship |
1 Sonnet 5: introductory price $2 / $10 until August 31, 2026, then $3 / $15. 2 Gemini 3.1 Pro: standard rate up to 200K context; above that $4 / $18. All figures per provider docs or price trackers (see Sources).
Two things stand out immediately. First: between the cheapest (Flash) and the most expensive output (GPT-5.5) there is more than a threefold difference — the choice of model is a real lever. Second, and more important: even the most expensive model usually costs only fractions of a cent per request. Anyone who concludes from this that AI is "practically free" is overlooking the other two levels. Those are exactly what we look at now.
The hidden costs — where the budget really goes
The token price is the tip of the iceberg. Below it lie six items that, in sum, almost always make up the lion's share.
Integration & development
Talking to a model over an interface (API) is easy. Connecting it usefully to your systems — inbox, CRM, inventory management, knowledge base — is the real effort. Designing and testing prompts, building interfaces, catching edge cases: that quickly adds up to several person-days or person-weeks. With an external service provider you land here, depending on scope, in the four- to five-figure range — one-time, but due before the first productive token flows.
Human review (human-in-the-loop)
For everything that goes out or makes decisions, you need a human to double-check — at least in the early phase. This review time is an ongoing cost block and often the largest of all: if an employee checks 100 AI drafts a month at five minutes each, that is more than eight working hours — month after month. The AI does the work faster; unsupervised, it does not do it.
Maintenance & operations
Models get retired, prices change, providers release new versions on a quarterly cadence. Prompts that worked yesterday deliver different results after a model change. Anyone running an AI system in production needs someone who maintains, monitors, and adapts it — not a one-time project, but a permanent line item.
Infrastructure & hosting
Around the model, software runs: a server for your application, a database, logging, monitoring, often a vector database for document search. Anyone hosting in the EU for GDPR reasons, or running a model in-house, pays noticeably more here than for the tokens alone. For most setups these are double-digit to low triple-digit amounts per month — predictable, but real.
Data & GDPR
AI is only as good as the data it works with. Collecting, cleaning, and providing that data in a legally sound way takes time. On top of that comes the compliance effort: data processing agreements, possibly a data protection impact assessment, labeling obligations. This rarely produces large invoices — but it costs lead time and care, and whoever skips it pays for it later.
Weiterlesen — kostenlos
Den vollständigen Inhalt freischalten
Trag deine E-Mail-Adresse ein und bestätige sie: Du abonnierst den Signal-Forge-Newsletter von FORGE und erhältst sofort Zugang zu diesem und allen weiteren registrierungspflichtigen Inhalten. Die Abmeldung ist jederzeit möglich.
Schon registriert? Der Link aus deiner Bestätigungs-Mail schaltet dieses Gerät wieder frei.
Error costs & ramp-up
Two items that produce no invoice and still count: first, error costs — a wrong AI answer that nobody checks can end up costing more than a whole year of tokens. Second, ramp-up: your team has to learn to work with the tool, to trust it, and to know its limits. Both belong honestly in the budget.
Two transparent worked examples
Enough theory — let's work through two typical cases. The token figures are deliberately assumed to be conservative and serve the order of magnitude, not the decimal place.
Example 1: pre-sorting 1,000 support requests per month
Assume an agent reads each incoming request (about 500 tokens of input) and outputs a category plus priority (about 50 tokens of output). On the budget tier (Gemini 3.5 Flash, $1.50 / $9) that costs: 0.5 million input tokens × $1.50 = $0.75, plus 0.05 million output tokens × $9 = $0.45 — around $1.20 per month. Even on Opus 4.8 it would be around $4. The tokens here are practically free. The budget sits in the one-time hookup to the inbox and in the person who, at first, spot-checks whether the sorting is correct.
Example 2: producing 100 quote drafts per month
Here the AI produces real text: per draft roughly 2,000 tokens of input (request + templates + product data) and 1,500 tokens of output. On a strong workhorse (Sonnet 5 at the standard price $3 / $15) that is 0.2 million input × $3 = $0.60 plus 0.15 million output × $15 = $2.25 — just under $3 per month. On Opus 4.8 it would be around $5. But before even one of these 100 drafts goes to a customer, a human reviews it — say five minutes per draft. That is a good eight working hours a month. The review time costs a multiple of the tokens, every single month.
The core lesson from both examples: The token price is the smallest and most predictable item. Anyone who wants to estimate AI costs realistically calculates the human review time and the one-time build first — and treats the tokens almost like a rounding error.
How to actually cut the costs
If the tokens barely matter, the savings lever lies elsewhere. Three principles deliver the most.
Tiered model routing
Don't use the most expensive model everywhere. Routine — classifying, tagging, pre-sorting — runs on the budget tier. The workhorse (Sonnet-5 class) handles most of it. You save the expensive top model (Opus 4.8) for the last few percentage points of precision. Which model is good for what we broke down in detail in the Model Comparison 2026. This tiered approach often cuts usage costs by more than half — with no noticeable loss of quality.
Use caching
When the same system prompt or the same knowledge base always sits up front, prompt caching reduces the cost of that part by up to around 90 percent. For agents with a fixed context, this is one of the biggest and most easily captured levers.
Start small, then scale
The most expensive mistake is building big before the value is proven. Take one recurring process, automate it cleanly, measure the result — and only then expand. That keeps the build costs manageable, and you don't pay for features nobody uses.
So what does AI really cost?
Honest answer: the tokens cost almost nothing — the integration into your everyday work costs money. An AI project that consumes a few euros of tokens a month can still carry a four- to five-figure total effort in the first year, once build, review, and operations are honestly counted. That is not bad news — it just means you have to plan in the right places.
That is exactly how we work at FORGE: we route in tiers by task instead of picking the most expensive model across the board, we cache fixed contexts, and we build every use case small and measurable first, before it scales. Anyone who wants to rebuild the concrete steps themselves will find them in our hands-on playbook on AI agents. Model prices will keep falling — the hidden costs remain. Whoever plans for them from the start makes the better decisions.
Sources
- Primary Anthropic — Claude Sonnet 5 (price $2/$10 → $3/$15, positioned cheaper than Opus/GPT-5.5/Gemini 3.1 Pro): anthropic.com/news/claude-sonnet-5
- Primary Claude Platform Docs — Opus 4.8 ($5/$25 per 1 million tokens; prompt caching approx. 90 % cheaper): platform.claude.com/docs
- Primary OpenAI API Docs — GPT-5.5 ($5/$30, cached input $0.50, ~1 million context): developers.openai.com
- OpenRouter — Google Gemini 3.1 Pro (price $2/$12; tier >200K: $4/$18): openrouter.ai/google/gemini-3.1-pro-preview
- OpenRouter — Google Gemini 3.5 Flash (price $1.50/$9): openrouter.ai/google/gemini-3.5-flash
- TechCrunch — Anthropic launches Claude Sonnet 5 (price comparison, positioning): techcrunch.com