WRITER's new flagship is a post-train of Z.ai's GLM-5.2, and that is the story

WRITER shipped Palmyra X6 on August 13 alongside a rebuilt agent harness and governance tooling, available to its clients the same day ^[1][16]. Palmyra X6 is a post-training variation of Z.ai's open-weight GLM-5.2, according to TechCrunch and The Next Web, with WRITER using that model as a starting point rather than training from scratch ^[5].

That sourcing decision tells you more about enterprise AI in 2025 than any of the accompanying percentages. An enterprise vendor's flagship is now a derivative of a Chinese lab's open-weight release, and the same product adds administrator controls for enabling Anthropic, OpenAI, or custom models through Microsoft Azure, Amazon Bedrock, or NVIDIA NIM ^[13]. Read those two facts together and the positioning is consistent: the defensible layer is orchestration and governance, not the weights.

The numbers around the launch are softer. WRITER reports that the combined model and harness cut average agent costs by 52%, improve speed by 48%, and raise quality by 10% against prior performance ^[2], with its model page listing an average finished-task cost of $0.12 ^[3]. That implies a prior-generation figure of roughly $0.25 per finished task ^[17]. All of it is vendor-reported from WRITER's own evaluations rather than independently verified benchmarking ^[4]. TechCrunch reports a looser version of the same estimate, up to 50% on basic tasks ^[8].

The internally consistent part is the harness. CMSWire reports that across the models WRITER tested, the rebuilt harness alone completed tasks 44% faster and at 41% lower cost on average ^[10], and TechCrunch cites a WRITER research paper putting harness-efficiency gains at an average 40% across tested models ^[9]. Compose the harness-only cost figure with the headline claim and the residual attributable to anything else, including the new model, is about 19% ^[18]. The model swap is not doing most of the work here.

WRITER's researchers say as much: "The harness is the one component whose efficiency multiplies across every model an organization runs-present and future" ^[11]. The Next Web describes the harness as the layer determining execution across planning, retrieval, tool calls, validation, and retries ^[6], which is where the cost actually accumulates, since repeated calls and growing context inflate token consumption even when the user sees a single final response ^[7]. The release also adds centralized reporting for adoption, spend, and Playbook and Skill performance ^[12].

Pricing is listed at $2 per million input tokens and $8 per million output tokens, a 4x output multiple ^[14][19], and WRITER says an agent can work on one objective for up to eight hours ^[14]. An eight-hour autonomous run is exactly the workload where a per-task average tells you least, because the trace, not the sticker price, sets the bill. CEO May Habib told TechCrunch, "I think the enterprise is absolutely sick of chasing the next benchmark. They want flattening cost, and it seems like nobody can deliver that" ^[15].

Three things to watch. First, whether buyers can reproduce the 41% harness figure on their own traces, particularly workloads with long context, retrieval, external tools, and human approval steps. Second, whether the spend reporting breaks out retries and context growth per task, or only totals. Third, whether other enterprise vendors start disclosing base-model provenance as plainly as this release did, because procurement questions about licensing and update cadence now point at a lab most of these buyers have never contracted with.