Writer sells AI tools to marketing teams and large enterprises like Accenture, Intuit, and Uber. On Thursday, it launched a new flagship model called Palmyra X6, alongside an upgraded version of the software that runs its AI agents. Together, Writer says the pair can cut customer costs by as much as half for routine tasks.
The interesting part is not the model's performance. It is what the model is built from. Palmyra X6 is not built from the ground up. It is a modified, retrained version of GLM-5.2, an open source AI model released by Z.ai, a Chinese AI company. Writer took that free foundation, tuned it for business use, and now sells it under its own brand name.
This is becoming a common playbook. Building a top AI model from scratch costs hundreds of millions of dollars. Taking a strong open source model and customizing it is far cheaper, and Chinese labs like Z.ai have been giving away increasingly capable models for free. For companies buying AI tools, this means the product marketed as "American enterprise AI" may quietly run on Chinese-built technology underneath. That is worth knowing, especially for companies in regulated industries like finance or healthcare, where your software actually comes from can matter for compliance.
The cost story is the bigger point. Writer published research testing what actually drives AI spending down: the choice of model, or the software that connects the model to real business tasks, often called the harness. Across six different AI models tested on the same 22 business tasks, changing the harness alone cut token costs by an average of 40 percent, while task success rates went up slightly rather than dropping. In plain terms, how you wire an AI system together to do a job matters more than which brand of AI powers it.
This lands at a moment when AI bills are becoming a real headache for businesses. Industry research shows the price of AI processing has fallen more than 300 times over in a few years, yet actual company spending on AI keeps rising sharply, because usage is growing even faster than prices are falling. Businesses are not paying less. They are paying more, for more.
Writer's CEO, May Habib, frames this as enterprises losing patience with the big AI labs, which she says have a financial incentive to keep usage climbing since that is how they earn revenue. That claim deserves a grain of salt, since Writer is also selling an alternative it profits from. But the underlying tension is real: when a vendor gets paid by the token, it has no natural reason to help you use fewer of them.
For any business now running AI agents to handle customer service, document review, or research tasks, the lesson is not to chase the newest model. It is to ask vendors for real cost-per-task numbers, and to question pricing built around volume rather than results.