Companies Reassess AI with Rising Tokenmaxxing Costs

Companies Seek Cheaper AI as 'Tokenmaxxing' Loses Momentum

The "tokenmaxxing" craze in corporations is cooling off. What seemed like a magic solution to boost productivity with artificial intelligence now faces a hurdle: rising costs. Companies that heavily invested in tools like OpenAI's ChatGPT and Anthropic's Claude are realizing that the return is not as high as they expected. The idea that more tokens mean more efficiency is falling apart.

Initially, Silicon Valley executives encouraged excessive token usage as a sign of performance. The typical "tokenmaxxer" was that employee who spent sleepless nights orchestrating a legion of AI agents to do the work for them. This generated a revenue wave for language model developers, but it has not proven to be a viable strategy for everyone. According to Jue Wang, a consultant at Bain & Company, many large companies are reassessing their investments in AI. The cost per token is doubling almost every month. Imagine spending $200 per developer per month and multiplying that by 20,000 developers. The result is a cost that no manager anticipated.

Not everything needs an advanced model like Claude Opus 4.6, which is more suited for software engineering or deep research. Many companies end up using these models for simple tasks, such as generating emails. This has led to the search for tools that perform "model routing," sending simpler queries to cheaper and more efficient AI systems, while complex tasks go to more powerful models.

On the other hand, those still betting on token accumulation are enjoying new open-source models from Chinese startups, such as Kimi from Moonshot or GLM from Zhipu. These models nearly match the capabilities of American models but cost significantly less. Raffi Krikorian, CTO of Mozilla, believes the industry is realizing that tokenmaxxing is a silly metric, similar to the old practice of measuring programmer productivity by the number of lines of code written. Just as this metric fell out of favor, tokenmaxxing seems to be following the same path.

The industry is waking up to reality. Tokenmaxxing may still have a bit more life thanks to Chinese models, but overall, it appears to be a passing trend. As Krikorian puts it, it’s a phenomenon that we will likely look back on and laugh at in a year.