Writer’s Palmyra X6: Cost-Slashing AI Agents Built on Open-Source GLM-5.2

Enterprise AI agents have promised big efficiency gains but often deliver bigger-than-expected bills. On August 13, 2026, Writer launched Palmyra X6, a new flagship model built as a post-training variation on Z.ai's open-source GLM-5.2 model . The company also released significant upgrades to its Writer Agent harness — the orchestration platform that controls how AI models execute complex, multi-step tasks . Writer’s pitch is direct: these changes can cut the cost of running sophisticated AI agents by more than half, potentially reshaping how large enterprises budget for AI automation.

What Writer launched: Palmyra X6 and a redesigned harness

The announcement had two core components:

• Palmyra X6: A new flagship model described as an agentic model purpose-built for long-horizon enterprise workflows. It supports sub-agents, grounding, and MCP tool use . Writer’s documentation lists it as the default model for the Writer Agent offering .
• Upgraded Writer Agent harness: The agent orchestration platform was redesigned to optimize token usage across long-running, multi-step agentic workflows . This harness controls how the model sequences reasoning steps, calls external tools, and manages context.

How harness optimization reduces costs

The cost savings claimed by Writer come not from a single magic trick but from a combination of model efficiency and smarter orchestration:

• 52% lower average cost: Agents paired with Palmyra X6 run complex multi-step workflows at an average of 52% lower cost compared to the previous generation .
• Per-task cost halved: The cost per agent task dropped from approximately $0.25 to approximately $0.12 .
• 48% improvement in speed: Execution time for agent tasks is roughly half of what it was before, with tasks completing in about 26 seconds at an output rate of 82 tokens per second .
• 10% better output quality, according to Writer’s internal benchmarks .
• Unattended long-running agents: The new harness and model can run agent jobs unattended for up to eight hours, reducing the overhead of manual supervision and re-prompting .
• Targeted token optimization: The harness directly optimizes how tokens are consumed during multi-step reasoning, cutting waste in long chains of agent actions .

Context: Writer specializes in AI tools and agents for marketing and revenue teams, serving Fortune 500 clients including Accenture, Uber, and Vanguard . The company’s core claim is that by pairing a frontier-competitive model with a smarter orchestration harness, enterprises can deploy persistent, multi-step AI agents at “a fraction of the cost” of competing solutions .

Why the open-source foundation matters

Palmyra X6 is based on Z.ai’s open-weight GLM-5.2, an approach that lets Writer benefit from an open-source base while adding proprietary post-training that tailors the model for enterprise agentic tasks . This hybrid strategy — taking a capable open model and fine-tuning it for domain-specific workflows — is increasingly common among AI companies looking to control costs and development time while still delivering competitive performance.

What this means for enterprise AI buyers

For enterprises evaluating AI agent platforms, Writer’s numbers suggest the economics of running complex, multi-step AI agents may be shifting. The combination of a 52% cost reduction, near-halved latency, and the ability to run agents unattended for eight hours addresses two of the most persistent complaints about enterprise AI agents: unpredictable costs and the need for constant human oversight. If these claims hold up in practice, they could make agentic AI viable for high-volume workloads that were previously too expensive to automate.

Of course, these figures come from Writer’s own internal testing, and real-world results may vary depending on the complexity and structure of specific enterprise tasks. Independent validation of the claimed cost reductions would provide a clearer picture for buyers.