Lovable spent $85,000 on LLM tokens over months of agentic coding experiments. The cost breakdown reveals that debugging often exceeds generation in token consumption.
Key facts
- $85,000 spent on LLM tokens for agentic coding experiments.
- Debugging accounts for ~60% of total token spend.
- Simple generation costs $0.50; complex debugging costs $15–$30.
- Long-context windows increase per-task token consumption.
- Agentic coding costs are not yet predictable for enterprises.
In a detailed post-mortem published on their engineering blog, Lovable disclosed spending $85,000 on LLM tokens across thousands of agentic coding sessions [According to $85,000 in tokens later](https://news.google.com/rss/articles/CBMihwFBVV95cUxPOHMwNzlIcHJyUnRBUW5RV0l3MHlUTkRibjc2elU2eTB1dVBJcFZxTWw3MDAtZVh5N3FDRW9XMFRVZ3RPMmhnX1NTbUwxNEhSRFI2VEp4ZDFLdjVzd0FqZzdjRmxmU1NlRHJTbUI4eUxfS1VBWXQ5c1FZME9lR1RURFFCVVdXcVU?oc=5]. The company did not specify which models they used, but the scale of spend offers a rare window into the real-world economics of agentic code generation.
Token costs scale non-linearly
The most striking finding: agentic coding costs scale non-linearly with task complexity. A simple function generation might cost $0.50 in tokens, but debugging a multi-file refactor can run $15–$30. The team noted that long-context windows, which many vendors tout as a feature, dramatically increased per-task token consumption because the agent re-reads entire files on each loop.
Debugging dominates the bill
Lovable's data shows that debugging and iteration cycles account for roughly 60% of total token spend. Initial code generation is cheap; fixing the generated code is where the cost compounds. This mirrors findings from other agentic coding startups that have privately shared similar ratios with us.
Implications for enterprise adoption
The unpredictable token bill poses a fundamental challenge for enterprise adoption of agentic coding tools. Finance teams accustomed to fixed-cost software licenses will struggle with variable costs that spike unpredictably when agents hit complex bugs. Lovable's experiments suggest that agentic coding is not yet cost-predictable for enterprises.
What to watch
Watch whether agentic coding platforms introduce fixed-price tiers or token caps in Q3 2026. If Lovable or competitors like Cursor or GitHub Copilot announce usage-based pricing reforms, it signals the market is maturing. Also watch for model providers like Google (which competes with OpenAI and Anthropic in this space) to release context-window optimization features that reduce token waste.
Source: news.google.com
Key Takeaways
- Lovable spent $85K on tokens for agentic coding.
- Debugging costs dominate, challenging enterprise adoption.
