AI in the Wild: What It Actually Costs to Build with It

Digit Conference, Tartu | October 9, 2026
Speaker: Ivan Elorza Tahum, CTO, Amiqa Tech AB
Written by Rene Rumberg

The Real Cost of Building with AI: A Practical Guide

Building with AI sounds exciting, but what does it actually cost? Ivan Elorza Tahum, CTO of Amiqa Tech AB, broke down the financial reality of working with large language models (LLMs) in a recent keynote.

Understanding Token Economics

The foundation of AI costs is simple: you pay for tokens. One token roughly equals one word or symbol. But here's where it gets interesting - different models have vastly different pricing structures.

Ivan compared several popular models: Claude's Haiku is cheap but limited, Sonnet offers a good middle ground at $2 per million input tokens and $10 per million output tokens, and Opus 5.5 is the most powerful but also the most expensive at $8 per million tokens.

The Hidden Costs of Context

Raw token pricing tells only part of the story. When you ask a question, the model loads instructions and tools into context - often adding hundreds of tokens to your bill automatically. Ivan demonstrated how a single question could cost $1 using Opus 5.5, largely due to context overhead.

But there's a bigger trap: as conversations grow longer, context windows expand, and costs skyrocket. Ivan showed an example where a 4-hour session ended up costing $320 due to accumulated context - nearly 15 times more than expected.

Strategies to Cut Costs

The good news? You don't have to accept these costs. Ivan shared several proven strategies:

  1. Use agent-based workflows: Deploy cheaper models (like Haiku) as workers for specific tasks, reserving expensive models only for complex reasoning. This hybrid approach maintains quality while dramatically reducing costs.

  2. Manage context carefully: Compress conversations, archive old context, and don't keep everything in memory. Proper context management can reduce costs by 70%.

  3. Leverage newer, more efficient models: Sonnet 5.5 matches Opus 5.5's quality at a fraction of the cost, with significantly cheaper token pricing.

  4. Parallelize work wisely: The team implemented rules like "agents can't spawn agents" and "waves of three to four" to manage load while controlling expenses.

Key Takeaway

AI isn't prohibitively expensive, it's just expensive in ways most developers don't expect at first. Understanding token economics, context management, and smart model selection can transform sky-high costs into manageable budgets. The teams winning with AI aren't necessarily using the most powerful models; they're using the right model for each job.