Writer Palmyra X6 Pairs Open-Weight GLM-5.2 With Cheaper Agent Harness

What Writer shipped

On August 13, the enterprise AI company Writer released Palmyra X6, its new flagship model, together with an overhauled agent harness. Writer pitched both as ways to rein in the token spending that has followed companies into production agent deployments.18 Writer started out selling AI tools and agents to marketers, and both releases went to its existing clients on launch day.1

Palmyra X6 was not trained from scratch. It is a post-trained variant of GLM-5.2, the open-weight mixture-of-experts model from Beijing-based Z.ai, which used to be called Zhipu AI, and Writer names that base model in its technical report.82 Around the model, Writer rebuilt the orchestration layer that controls how agents plan, batch work, hand tasks to sub-agents and recover from failed tool calls.7 It also added governance tooling meant to let IT administrators see and limit token spending.8

Outlets reported the headline numbers in slightly different ways. TechCrunch said the model and harness together could cut costs by as much as 50% on basic tasks.1 VentureBeat and SiliconANGLE used Writer's average figures: 52% lower cost, 48% faster and 10% better quality when Writer Agent runs on Palmyra X6.89 The two framings are compatible, since one describes an upper bound on simple work and the other an average across tasks. Both come from Writer's own measurements, and no independent testing has confirmed them.2

The harness is the real story

For developers, the most important part of the launch is the claim that orchestration code can save more money than switching models. A Writer research paper tested small harness-efficiency changes across several models. It found that harness changes cut costs more reliably than model choice did, with an average reduction of about 40%.1 Other accounts gave more precise numbers. One reported that the harness alone made Writer Agent 44% faster and 41% cheaper per task across every model tested, including third-party ones.7 Another cited the paper's figures showing tokens per task falling from 14,200 to 8,800 and cost per task from 21 cents to 12 cents, with the same models running the same tasks.10

The researchers argued that the harness is the one component whose efficiency gains apply across every model an organization runs, now and in the future.1 That is a strong claim, and it fits a wider shift in how practitioners think about agent costs. An independent analysis published in October measured the context each coding harness loads on its first call. It found roughly 2,000 tokens for Pi, about 11,300 for Codex and about 27,000 for Claude Code. On SWE-bench Lite, a single attempt cost anywhere from 3 cents to $1.54 depending on the model and app.15 A September arXiv paper on enterprise coding agents made a related point. Most enterprises buy harnesses as products, and the harness picks the model, writes the prompts, manages the cache and launches subagents. So while the price sheet sets what a token costs, the harness decides which rate applies and how many tokens get bought at it.13

Other teams report similar gains from changing the harness alone. One team building on a different tool described moving repetitive agent steps out of LLM calls and into code. In their account, per-alert model costs on a compliance workload fell from $2.89 to $0.25.12 That is a separate product and a different workload, so the numbers can't be compared directly with Writer's. Still, it points the same way: a large share of agent spend comes from orchestration overhead, not from the model's per-token price.

The Writer paper does include an important caveat. Sub-agent delegation, one of the harness's main cost-saving techniques, worked reliably only on strong models. In Writer's testing, only Palmyra X6 (0.86 reliability) and Claude Sonnet 4.6 (0.85) cleared the usable threshold.10 So the harness savings do not carry over to small, cheap models as well as the headline numbers suggest.

How X6 was built on open weights

The training approach will interest anyone working with open-weight models. According to Writer's technical report, X6 keeps GLM-5.2's architecture unchanged: 744 billion total parameters, about 40 billion active per token.8 Writer's post-training was deliberately conservative. It used a method it calls anchored supervised fine-tuning on just 626 curated synthetic agentic trajectories, trained for a single epoch at a low learning rate.8 The method pairs token weighting with a KL-divergence anchor, which penalizes the model for drifting away from a frozen copy of the base. The goal is to teach new tool-use behavior without wearing down the base model's general abilities. Writer also replaced the Adam optimizer with Muon on the model's core weight matrices.8

All of the training data was synthetic. Teacher models generated it, and it passed through structural checks, a model-based verifier and a two-model LLM judging panel before use.8 Dan Bikel, who leads Writer's AI research, called X6 "very much a Palmyra model", with the GLM weights serving only as a starting point.2

The term "open source" needs care here. Several reports describe GLM-5.2 as open source.19 But Palmyra X6 itself is not open. Writer sells it through its own API and platform and has not released the weights, so the post-trained model is proprietary even though its base is open-weight.7 Developers looking for a downloadable checkpoint won't get one. What Writer has published is a fairly detailed recipe for adapting an open base model with very little data, and many teams can reproduce that approach on their own.

Pricing, performance and the caveats

VentureBeat reported that Writer prices X6 at $2 per million input tokens and $8 per million output tokens, compared with $15 and $75 for Anthropic's Claude Opus 4.8.8 One aggregator said it could not confirm the pricing from the sources it checked.2 For comparison, Amazon's Bedrock pricing page lists the previous Palmyra X5 at $0.003 per thousand input tokens and $0.015 per thousand output tokens.16 If the X6 figures hold, the new flagship costs less per token than the model it replaces.

On Writer's internal evaluation of nine agent capabilities, X6 averaged 0.87. That put it ahead of Claude Opus 4.8 at 0.86, GPT-5.5 at 0.80 and Gemini 3.1 at 0.77.7 Writer chose both the test and the task mix, and the scores are best treated as a sign of fit for marketing and revenue workflows, not as evidence of frontier-level performance overall.72 Writer also says X6 completes tasks in 26 seconds on average, generates 82 tokens per second and can work toward one goal unattended for up to eight hours.9 It has a context window of 1 million tokens.7

The provenance question

This is where the reporting differs most. VentureBeat put the most weight on the risks of building on Chinese open-weight models. It cited a SaferAI report finding that GLM-5.2 refused none of the offensive cyber or biology tasks it was given through Z.ai's public API, and that Z.ai had published no safety framework.8 Writer responds that provenance and post-training matter more than where a model comes from. Bikel said the weights were downloaded from the US Hugging Face and that all training ran on US hardware. Writer also reported that X6, with its deployment system message, scored 8.6 points higher on the FORTRESS adversarial safety benchmark than the raw base model.8 Other outlets mostly treated GLM-5.2 as a smart, low-cost starting point that fits the efficiency pitch.310 One noted that whether legal and security teams in regulated industries accept Writer's framing is a separate question from whether it is technically accurate.2

Our view: for most developers the supply-chain question is real but manageable. Writer has been unusually open about where X6 came from, and that openness is what makes a proper review possible. Teams with strict procurement rules should still check that review themselves and not rely on the vendor's benchmark results.

Why it matters for developers

The most practical part of the launch may be the governance tooling. Writer's AI Studio now offers live token reporting, alerts and hard blocks once spending passes a set limit.7 Writer Agent also gained multi-model support, so admins can enable models from Anthropic, OpenAI, Microsoft Azure, AWS Bedrock and Nvidia NIM.8 CTO Waseem AlShikh summed up the goal: enterprises want token consumption to grow because that means adoption, but they need costs to level off.9 Asked why Writer builds its own model when the harness provides most of the savings, he said the company cannot control when a lab deprecates its model.10 CEO May Habib also pointed to growing distrust of the big AI labs, which she said have a financial incentive to push token usage higher.5

Overall, the launch shows open-weight models becoming a commodity starting point, with the real competition moving to post-training and orchestration. Writer's 50% figure is a vendor claim. But the claim that harness design matters as much as the model has support beyond Writer's own paper.151312 For developers, the main lesson is to measure how many tokens their harness spends before switching to a different model.