top of page
Search

What Does AI Training Actually Cost? Real Invoices From Running Live AI Learning Tools

11 minutes ago
5 min read

Every conversation about AI in training eventually reaches the money question, usually phrased carefully by someone who has to justify a number to a finance officer. Demonstrations are everywhere. Published invoices are not. So I pulled my own billing for the two live AI tools I run: an instructional design coaching tool on this website, built on the Claude API, and a traffic stop scenario where the learner talks out loud to an AI driver, built on ElevenLabs voice agents.

The usage was trivially cheap. The structure of the purchases was not. This post walks through the real AI training costs I found, the three different kinds of bills training developers tend to confuse, the factor that actually picks your pricing plan, and four steps for budgeting AI-powered training before you put it in front of learners.

The Per-Learner Cost of AI Training Is Smaller Than You Think

The coaching tool cost twenty-two cents for the month from October 5 to November 3: twelve cents of input, ten cents of output, across six days of activity. Since April it has consumed four dollars and fifty-six cents in total.

The traffic stop scenario runs up to about eight minutes per session, including the debrief. ElevenLabs charges eight cents a minute for agent conversations, and that rate is the same on every tier I checked. Eight minutes at eight cents is sixty-four cents per learner. A hundred officers costs sixty-four dollars; a thousand costs six hundred and forty.

One caution: that figure covers the voice. The language model behind the agent is billed separately, so sixty-four cents is a floor, not a total.

Three AI Bills Training Teams Confuse

There are three different purchases here, and they behave differently.

  • A subscription. I pay $20 a month for Claude, the chat product I work in. It is a flat fee for one person at a keyboard, and I have never exceeded it.

  • API access. An API key is a long credential that lets your own website or tool use a company's AI, with the usage billed to you. That is the twenty-two cents. Paying for the subscription gives you nothing on the API, and anyone holding your key can spend your money, so never embed one in anything a learner can download.

  • A voice platform subscription. ElevenLabs bills conversations by the minute against a monthly bucket with a cap on simultaneous sessions, while images, video and voiceover draw on a separate credit allowance. Running out of one does not touch the other.

Put together, my whole operation runs about $315 a year. Three quarters of that is the subscription I think in. The learner-facing infrastructure, the part that feels expensive and futuristic, is under eighty dollars a year combined.

The Hidden Costs Are Structural, Not Usage

The largest line item in this entire project was money I spent on nothing. I bought $21 of Claude API credits in April because that was the smallest amount I could buy. Anthropic's own help centre confirms that prepaid credits expire one year from the purchase date and are non-refundable. At my current rate, roughly $13 to $15 of that $21 will expire unused.

Production months and running months are the second structural cost. A seventeen-second video lead-in for the traffic stop scenario consumed roughly two thirds of a thirty-thousand-credit monthly allowance. I upgraded for that one month to finish it and moved back down once production ended. Most teams buy one plan for the year and pay production prices through months of light use.

This matters because budgets are tight. Fosway's Digital Learning Realities 2026 research found that one in four L&D teams had budget cuts this year, while 63% expect spending on AI content authoring to grow. Every expired credit and oversized plan gets harder to defend.

Concurrency, Not Volume, Picks Your AI Pricing Plan

Your plan is not chosen by how many minutes you use. It is chosen by how many people press start at the same moment. On ElevenLabs, the $22 tier allows ten simultaneous conversations, the $99 tier allows twenty, and thirty puts you on the $299 plan. Go over the limit and burst pricing charges excess calls at double the standard rate.

Picture a police academy. Thirty recruits in a classroom all start the scenario when the instructor says go. That needs the $299 plan. The same thirty officers working through it individually over a fortnight fit comfortably on the $22 plan. Same training, same minutes, more than ten times the subscription cost.

The useful side effect: spaced, individual completion, which I would argue for on retention grounds anyway, is also the pattern that keeps you on the cheap tier.

Model Choice and the Free-Tier Pilot Trap

Every token my coaching tool used was Haiku, Anthropic's small, inexpensive model, not the flagship. Structured feedback against a rubric did not need the most powerful model available. Anthropic currently lists its newest small model at ten cents per million input tokens, a fraction of flagship pricing. If a costing alarms you, check which model you specified before deciding the idea is unaffordable.

The free tier creates a different problem. ElevenLabs includes fifteen minutes of agent conversation on its free plan, which is one or two learners at eight minutes a session. That is enough to prove the technology works and useless for proving it can be operated. A pilot that proves capability is not a pilot that proves viability. UC Today's reporting on AI pilot purgatory notes that pilot ROI models usually count licence cost and labour saved while leaving governance and operating costs out entirely.

How to Budget AI-Powered Training: Four Steps

  • Work out concurrency first. How many learners will be in the same scenario at the same moment is the question that selects your plan, and your existing timetable usually answers it.

  • Separate build months from running months. Resist annual billing until you know what a quiet month costs, because annual plans are sized for your busiest week.

  • Check minimums and expiry on prepaid credits. Mine cost me more than my actual usage did.

  • Test on the small model first. You can always move up if the output is not good enough.

A caution on my figures: I have had tens of users, not thousands, on consumer-tier accounts. And the design time behind the traffic stop scenario dwarfed every subscription combined. The honest pitch for AI scenario training is not that it is cheap. It is that the marginal cost of one more learner is close to nothing, which only pays off if the scenario runs many times. The only numbers that will hold up in your business case are the ones you pulled yourself.

Want to Build These Skills With Your Team?

If you want your training team to start planning and budgeting AI-powered training tools with confidence, I offer private 4-hour virtual workshops designed specifically for training developers. We work through your department's actual projects using multiple AI platforms, so participants leave with practical skills and working materials, not just theory.

Format: Private virtual sessions for up to 20 participants

Investment: $2,000 USD / $2,500 CDN per workshop

Sources

A Note on AI Use

This post was researched and drafted with AI assistance, then reviewed and edited for accuracy and voice. All practical recommendations reflect my own instructional design experience.

 
 
 

Comments


bottom of page