Writer targets token costs with Palmyra X6 and harness overhaul

Writer isn't chasing another leaderboard spot with its newest model. It's chasing your invoice.

The enterprise AI company released Palmyra X6, its new flagship model, on Thursday alongside a substantial upgrade to its agentic harness, the infrastructure layer that governs how its models plan, call tools, and execute multi-step tasks. Both went live for customers the same day.

Rather than pitching Palmyra X6 as a benchmark leader, Writer is framing the release around a more mundane but pressing enterprise concern: the runaway token costs that come with scaling agentic workloads. CEO May Habib said enterprise buyers are now more focused on flattening those costs than on chasing raw performance gains.

"Enterprise buyers are focused on flattening costs, not chasing benchmarks," Habib said of the shift driving the release.

Writer says the combination of the new model and the retooled harness could cut costs for basic tasks by as much as 50%. That figure comes from the company's own estimate, and it hasn't been independently verified against competing agent stacks.

The harness improvements alone appear to carry most of the weight. Writer published internal research showing that efficiency changes to the harness, independent of which underlying model it ran, reduced costs by an average of 40% across multiple models tested. That suggests the savings are less about Palmyra X6's raw efficiency and more about smarter orchestration: fewer wasted tool calls, tighter context management, and reduced redundant token consumption during agent execution.

Notably, Palmyra X6 isn't built from scratch. It's a post-trained variant of GLM-5.2, the open source model from Chinese AI lab Z.ai, a detail Writer disclosed rather than obscured. Building on an open weights base lets Writer skip the enormous pretraining costs of foundation model development and instead focus its resources on fine-tuning and the harness layer, where it argues the real enterprise value now sits.

Writer is also positioning the release as agnostic rather than closed. Palmyra X6 can run alongside the company's other proprietary models or alongside third-party models imported via Azure or Amazon Bedrock, letting enterprise customers mix models within the same harness rather than committing exclusively to Writer's stack.

The bet underlying the launch is that as agentic AI moves from pilot projects into production, the cost of running these systems at scale, not their benchmark scores, will decide which vendors enterprises keep using.