On August 13, 2026, Writer launched its new flagship Palmyra X6 model for its enterprise-focused AI agent platform, alongside a rebuilt task orchestration layer, Writer Agent harness, and governance tools for monitoring usage and setting spending limits. The company says that running its product with Palmyra X6 reduces costs by an average of 52%, while improving speed by 48% and quality by 10%.
The move comes at a time when agent costs have become a major factor in enterprise AI technology purchasing decisions. Agents typically do not simply produce a single answer; instead, they perform repeated cycles of planning, retrieval, tool calling, verification, and retrying, which means they consume more tokens than traditional chatbots. According to Goldman Sachs projections cited in the article, token consumption could increase 24-fold between 2026 and 2030, reaching 120 quadrillion tokens per month, driven by the spread of continuously operating agents.
A model built on open weights
Writer does not claim to have trained Palmyra X6 from scratch. The model is a post-trained version of GLM-5.2, an open-weight mixture-of-experts model from the Chinese company Z.ai, formerly known as Zhipu AI. According to the company’s technical report, X6 retains the GLM-5.2 architecture and contains 744 billion parameters, with approximately 40 billion parameters activated per token.
Writer used a method it called anchored supervised fine-tuning, or ASFT, on a small dataset of 626 selected synthetic trajectories for agent tasks, for one training cycle and at a low learning rate. The method uses token weighting and a Kullback–Leibler divergence-based anchor to limit the trained model’s departure from a fixed version of the base model. The company also replaced the Adam optimizer with the Muon optimizer in the core weight matrices.
The training data was entirely synthetic: teacher models generated task plans, tool calls, and final answers, which then passed through structural quality gates, a model-based checker, and a jury composed of two language models before being used for training.
Pricing and performance according to Writer’s tests
In internal evaluations covering nine capabilities, including source-grounded retrieval, tool use, content generation, delegation of tasks to sub-agents, and adherence to brand voice, Palmyra X6 achieved an average score of 0.87 out of 1.00. Writer compares this result with 0.86 for Claude Opus 4.8, 0.85 for Claude Sonnet 4.6, 0.80 for GPT-5.5, and 0.77 for Gemini 3.1.
The company prices the model at $2 per million input tokens and $8 per million output tokens, compared with $15 and $75, respectively, for Claude Opus 4.8. It says X6 completes tasks in an average of 26 seconds and can operate without direct supervision toward a single goal for up to eight hours.
However, these comparisons are based in part on internal tests conducted by Writer itself. Dan Bikel, who leads the company’s AI research, said that the technical report includes the public testing protocol and internal evaluations, explaining that public tests are used as indicators to verify that there are no major gaps, rather than as the primary training targets.
Infrastructure and governance
Writer says the model weights were obtained from the Hugging Face repository in the United States, and that training was conducted entirely on U.S. infrastructure, with datasets generated and stored in the United States and training hardware located there. The company also conducted a preregistered assessment of model risks covering political bias, censorship, accuracy, and refusal behavior, involving 19,674 responses evaluated by anonymized judges.
Writer’s results indicate that X6 presented a side of sensitive political questions in 80% of cases in The Washington Post’s ModelSlant evaluation. It also achieved, when using the system message intended for deployment, an 8.6-point improvement in attack safety on the FORTRESS benchmark compared with the base GLM-5.2 version. However, the report acknowledged that the model’s behavior varies by language, meaning that limited post-training does not necessarily remove all effects of the base model.
Alongside the model, Writer rebuilt the orchestration layer that plans tasks, groups and delegates work to sub-agents, and manages context. The company says this layer reduces costs by 41% and accelerates task completion by 44% across all the models it tested, including Anthropic and OpenAI models, while maintaining quality.
Writer also expanded multi-model support within Writer Agent, enabling administrators to activate models from Anthropic and OpenAI, as well as cloud service providers including Microsoft Azure, AWS Bedrock, and Nvidia NIM, in addition to image-generation models.
A unified view of agent costs
The new governance tools add a centralized view of agent usage across the enterprise and analytics for each workflow within the shareable automation tools Playbooks and Skills, along with consumption alerts and spending limits. These tools are intended to help IT and security administrators understand the cost of each workflow before expanding its use.
Writer says the cost reduction is not intended to reduce its revenue, but to open use cases that organizations previously would not have automated because of labor costs or token consumption. The company presents the launch of Palmyra X6, the orchestration layer, and the governance tools as a unified system focused on the cost of a successful task, rather than solely on the advertised token price.