Writer introduces new AI model and upgraded harness to contain token costs

San Francisco-based AI company Writer, a leading enterprise AI agent platform, has announced a significant leap in making advanced AI agents more affordable and efficient for large organizations. The company recently unveiled its new flagship model, Palmyra X6, alongside major upgrades to its proprietary Agent harness. This dual release is specifically designed to tackle the escalating token costs associated with complex AI workflows, promising deployment-ready capabilities at a much lower price point.

Writer's Bold Step Towards Economical Enterprise AI

On August 13, 2026, Writer introduced Palmyra X6, a powerful new AI model, and a significantly improved "harness" for its AI Agent platform. This announcement marks a pivotal moment for enterprises grappling with the economics of scaling AI, particularly for agentic workflows that demand extensive token consumption. Writer's co-founder and CTO, Waseem AlShikh, emphasized the enterprise's desire for increased token consumption—a sign of adoption—but stressed the critical need for costs to remain flat.

Writer, known for providing a full-stack generative AI platform to Fortune 500 companies, has positioned itself as a solution for high-stakes enterprise environments. Their platform empowers marketing, sales, and business teams with AI teammates capable of planning, executing, and scaling on-brand work across company systems, all while embedding rich organizational context.

Palmyra X6: A Strategic Evolution from Open-Source Foundations

The Palmyra X6 is not built from the ground up but rather as a sophisticated post-training variation of GLM-5.2, an open-weight mixture-of-experts model developed by Beijing-based Z.ai (formerly Zhipu AI). This strategic choice is noteworthy, as GLM-5.2, released in June 2026 under the permissive MIT license, has quickly gained recognition as one of the most capable openly available models globally, particularly excelling in coding and long-horizon tasks with its solid 1M-token context.

Writer's approach involved a "deliberately conservative post-training recipe" using a technique called anchored supervised fine-tuning (ASFT). This was applied to a remarkably small corpus of just 626 curated synthetic agentic trajectories, trained for a single epoch at a low learning rate. This method allows Writer to leverage the robust capabilities of GLM-5.2 while fine-tuning it specifically for enterprise-grade agentic work.

Addressing potential concerns regarding the use of a Chinese open-source foundation, Writer explicitly states that Palmyra X6 is entirely disconnected from its original developers and runs fully on Writer's U.S. infrastructure. This underscores the company's commitment to security and trust for its enterprise clients.

The "Harness Effect": Orchestration as a Cost-Saving Powerhouse

Perhaps the most impactful aspect of this release is the significant upgrade to the Writer Agent harness. The harness is the intelligent orchestration layer responsible for planning tasks, batching work, delegating to sub-agents, and managing context within complex AI workflows. Writer's research, published in an accompanying paper titled "The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI," highlights its profound impact.

The company's research demonstrates that these harness upgrades alone can improve performance regardless of the underlying model. Across various Writer and third-party models tested, the Writer Agent harness enabled tasks to be completed 44% faster and at a 41% lower cost per task on average, all while maintaining quality. This finding suggests that optimizing the orchestration layer can yield greater cost efficiencies than simply swapping out models for cheaper alternatives.

When the upgraded harness is paired with the new Palmyra X6 model, the combined performance is even more impressive. Writer reports an average of 52% lower operational costs, a 48% improvement in speed, and a 10% boost in quality for its agent product. This represents a substantial reduction in the total cost of ownership for enterprise AI deployments.

Performance and Pricing Benchmarks

Palmyra X6 is not just cost-effective; it also delivers top-tier performance. On Writer's internal evaluations, which cover nine capabilities including grounding and retrieval, tool use, content generation, and brand voice, Palmyra X6 scored an average of 0.87 out of 1.00. This score puts it ahead of several prominent models in the industry, including Anthropic's Claude Opus 4.8 (0.86), Claude Sonnet 4.6 (0.85), OpenAI's GPT-5.5 (0.80), and Google's Gemini 3.1 (0.77).

The model is engineered for efficiency, completing tasks in an average of 26 seconds and capable of working unattended toward a single goal for up to eight hours.

From a pricing perspective, Writer has made Palmyra X6 highly competitive. The model is priced at $2 per million input tokens and $8 per million output tokens. This is a significant differentiator when compared to more expensive alternatives, such as Claude Opus 4.8, which is priced at 15/75 for input/output tokens respectively. This aggressive pricing strategy, combined with the harness's efficiency gains, aims to make large-scale agentic AI deployments economically viable for enterprises.

Industry Implications and Writer's Vision

The release of Palmyra X6 and the upgraded harness addresses a critical challenge in the rapidly evolving AI landscape: the rising cost of token consumption as AI agents move into production. As agents perform complex, multi-step workflows, call tools, and execute longer-running tasks, each action consumes more tokens, leading to substantial cost increases compared to traditional chat experiences.

Writer's focus on cost-efficiency, reliability, and governance reflects a deeper understanding of enterprise needs. By offering a full-stack platform that includes proprietary models, a Knowledge Graph, and robust governance tools, Writer enables IT and business stakeholders to build and supervise AI agents for complex, cross-functional workflows.

This development is set to accelerate the adoption of agentic AI within large organizations, allowing them to confidently scale ambitious AI workflows without token costs becoming a prohibitive factor. Writer's multi-model support also ensures that teams can choose from a selection of popular models, including their own and third-party options, preventing vendor lock-in and offering flexibility.

This move solidifies Writer's position as a leader in enterprise AI, demonstrating that innovation in orchestration and model optimization can fundamentally change the economics of AI deployment, paving the way for broader and bolder applications of agentic AI across industries.

Frequently Asked Questions

What is Palmyra X6?

Palmyra X6 is Writer's new flagship AI model, released on August 13, 2026. It is a post-trained version of Z.ai's open-source GLM-5.2 model, specifically optimized for enterprise AI agent workflows to deliver high performance at a significantly reduced cost.

How does Writer's new system reduce token costs?

Writer reduces token costs through two main innovations: the Palmyra X6 model itself, which is priced competitively, and major upgrades to its proprietary Agent harness. The harness is an orchestration layer that optimizes how AI agents plan, execute, and manage tasks, leading to more efficient token consumption regardless of the underlying model.

What are the performance improvements of the new system?

When Palmyra X6 is paired with the upgraded Writer Agent harness, the platform achieves an average of 52% lower costs, 48% faster execution speed, and 10% better quality compared to previous iterations. The harness alone can reduce costs by 41% and increase speed by 44% across various models.

Is GLM-5.2, the base model for Palmyra X6, safe for enterprise use given its origin?

Writer acknowledges that GLM-5.2 originates from Beijing-based Z.ai but emphasizes that Palmyra X6 is a post-trained version fully run on Writer's U.S. infrastructure. Writer states that the model is in no way connected to its original developers and that its provenance and post-training are more important than its initial origin for enterprise security and trust.