Writer Palmyra X6 launch pairs new model with cost-focused AI harness
Writer launched Palmyra X6 and an upgraded agentic harness, saying the combination can cut basic-task costs by up to 50%.
By Jordan Bell · Startups & Deals Reporter
· 3 min read
The Writer Palmyra X6 launch pairs a new flagship AI model with upgrades to the company’s agentic harness, software designed to control how an AI system carries out multi-step work. Writer said the combination could lower customers’ costs by as much as 50% on basic tasks, though that figure is the company’s estimate rather than an independently verified result.
Both products became available to Writer clients on Thursday, according to TechCrunch. For investors following enterprise AI, the announcement puts attention on a less visible part of the technology stack: not only which model a company uses, but how the surrounding software manages its requests, data and tools.
Palmyra X6 is a post-training variation of Z.ai’s open-source GLM-5.2 model, not a model Writer trained from scratch, TechCrunch reported. Writer plans to offer it alongside its other models and third-party models that customers import through Azure or Amazon Bedrock.
What is Writer’s AI harness and why does it affect costs?
A harness is the orchestration layer around a language model. It decides what context to send to the model, which tools it can use, when it should retrieve information, whether work should be delegated, and when the system should stop or retry a task, according to a July preprint by Writer employees.
That matters because an AI agent can make several model calls to finish one job. Repeating long histories, tool descriptions and retrieved documents can increase the number of tokens processed. Since providers generally charge based on tokens, reducing unnecessary material or repeated steps can lower the cost of completing a task even when the underlying model remains the same.
Writer’s product estimate covers the combined effect of Palmyra X6 and its harness-infrastructure changes on basic tasks. The company did not provide pricing, customer adoption figures or independently validated savings in the information reported by TechCrunch.
What did Writer’s research find?
Writer employees tested their Agent Harness in a company-authored arXiv preprint published in July. The study compared a conventional production agent loop, frozen on June 7, with Writer’s harness while keeping the prompts, tasks, model identifiers, judges and price tables the same.
Across 22 enterprise tasks and six models, the paper reported that blended cost per task declined 41%, from $0.21 to $0.12. Tokens per task fell 38%, from 14,200 to 8,800, while median completion time dropped 44%, from 48 seconds to 27 seconds.
The models tested were Claude Sonnet 4.6, Gemini 3.1, Gemini Flash 3.5, Qwen 3.6, GLM 5.1 and Palmyra X6. The paper reported task-completion quality rising from 0.78 to 0.81, but described that result as directional at the study’s sample size, rather than conclusive evidence of better quality.
The research provides a specific case for harness design as a cost lever, but its findings are limited to Writer’s selected tasks and experimental setup. Because the preprint was written by Writer employees, readers should view it as company research rather than independent validation of the launch claims.
This story draws on original reporting from TechCrunch.
