Palmyra X6: Writer Pairs GLM-5.2 Model With Cheaper Agent Harness

What Writer shipped

On August 13, 2026, the enterprise AI company Writer released Palmyra X6, its new flagship model. It shipped the model alongside a rebuilt agent harness and new spending controls, and the company's pitch is about the economics of running agents rather than raw intelligence.13 All of the new features became available to customers the same day.34 Writer sells its software to marketing, sales and revenue teams at large companies.3

The headline figure is that Writer's Agent product now costs 52% less to run on average when paired with Palmyra X6. Writer also says it runs 48% faster and scores 10% higher on quality.13 Co-founder and CTO Waseem AlShikh summed up the strategy this way: enterprises want token consumption to explode because that means people are using the tools, but they need costs to flatten.13

The release has three parts. There is a model built on an open-weight base, a harness the company argues matters more than the model, and a governance layer that puts a price on each workflow. Taken together, they show where the enterprise agent market is heading.

An open-weight Chinese base under an American enterprise brand

The most notable detail is where the model comes from. Palmyra X6 was not trained from scratch. It is a post-trained version of GLM-5.2, the open-source mixture-of-experts model from Z.ai.14 According to Writer's technical report, X6 keeps GLM-5.2's architecture unchanged: 744 billion total parameters, with roughly 40 billion active per token.9

Writer's own contribution was modest by design. The company used a technique it calls anchored supervised fine-tuning on only 626 curated synthetic agentic trajectories, trained for one epoch at a low learning rate.9 A KL-divergence anchor penalizes the fine-tuned model for drifting too far from a frozen copy of the base. The aim is to teach tool-use behaviour without degrading the capabilities GLM-5.2 already has.9 This fits Writer's history of cheap training: Palmyra X 004 was reportedly trained for about $700,000 in 2024, and X5 took roughly $1 million in GPU hours.9

The sources mostly agree on these facts but frame them differently. TechCrunch treats the GLM-5.2 base as a way to get the lower per-token prices of open models into a product that is ready to deploy.4 Coverage based on VentureBeat's interview treats it as a statement in a heated debate over whether American enterprises should build on Chinese open-source foundations.9 Writer's executives were clearly ready for that question. Product director Matan-Paul Shetrit said the model has no connection to its original developers and runs on U.S. infrastructure. AI research lead Dan Bikel said the team simply took the weights as a starting point and trained from there.9 Writer also says the weights were downloaded from Hugging Face's U.S. presence and that all training data and hardware stayed in the U.S.9

The choice says a lot about the open-weight market. GLM-5.2 came out in June under the permissive MIT license. It reportedly scored 51 on Artificial Analysis's Intelligence Index, ahead of DeepSeek V4 Pro, Kimi K2.6 and some of Google's Gemini models on agentic tasks.9 Z.ai has already moved on: one outlet covered GLM-5.3 the day after Writer's launch, describing China's open-weight coding race as speeding up.2 Writer has built its flagship on a base that its maker has already superseded. That is a risk, but also a repeatable process. If 626 trajectories is enough, Writer can probably re-anchor on each new open release cheaply.

The benchmark table, and where Gemini lands

Writer tested X6 on nine internal evaluations tied to production work, including grounding and retrieval, tool use, content generation, sub-agent delegation and brand voice.3 X6 averaged 0.87. That put it narrowly ahead of Claude Opus 4.8 at 0.86 and Claude Sonnet 4.6 at 0.85, and clearly ahead of GPT-5.5 at 0.80 and Gemini 3.1 at 0.77.3 X6 is priced at $2 per million input tokens and $8 per million output tokens.12 Writer's model page puts the average task cost at about $0.12.2

Gemini 3.1 came last in this comparison, at $2.50 input and $10 output per million tokens.13 That makes it slightly more expensive than X6 and the lowest scorer. Readers should treat this result with caution. These are Writer's own evaluations, tuned to brand voice and marketing workflows where Writer has trained specifically, and no independent replication has been reported. The result shows that X6 is competitive on Writer's home ground. It does not show that a lightly post-trained open model beats Google's general-purpose flagship.

The coverage also disagrees on what competitors charge. SiliconANGLE lists Claude Opus 4.8 at $3/$15.1 IT Brief lists it at $15/$75.3 Anthropic's own price page shows Opus 4.8 at $5/$25.15 None of these figures change the basic point, since X6 is cheaper than every one of them. But the spread shows that the price comparisons in launch materials were not checked carefully. The coverage also disagrees on Palmyra X5's age. SiliconANGLE dates X5 to April 2025,1 while model directories list a January 2026 release, at $0.60/$6 with a 1-million-token context window.810

Writer also says X6 is the least politically biased model it tested.1 On the Washington Post's ModelSlant evaluation, X6 reportedly presented both sides of hot-button questions 80% of the time, and it answered sensitive prompts that DeepSeek V4 refused.9 Given the model's origin, this claim seems aimed at buyers worried that the Chinese base model carries over its censorship.

On speed and endurance, Writer says X6 completes tasks in 26 seconds on average, generates 82 tokens per second, and can work toward a single goal unattended for up to eight hours.13

The harness is the real product

The model gets the headlines, but the more lasting claim is about the harness. Writer rebuilt the agent backend to scale reasoning to the task. Simple questions get direct answers, and complex work gets structured plans. The harness also adds batch execution and sub-agents for high-volume jobs.3 The goal is to cut unnecessary steps and avoid re-sending the same context, which raises costs and makes workflows more fragile.3

Writer says the rebuilt harness alone delivered 41% lower cost per task and 44% faster completion with no loss in quality. It says this held across every model it tested, including Anthropic's and OpenAI's.39 A Writer research paper on what it calls "The Harness Effect" found that improving the harness was often a more reliable way to cut costs than switching models, with costs falling about 40% on average.49 The researchers argue that the harness is the only component whose efficiency gains carry over to every model an organization runs.4

The percentages are not all measuring the same thing. The 52% figure applies to the harness paired with X6.1 The 41% figure applies to the harness alone.3 TechCrunch reports savings of "as much as 50%" for basic tasks.4 Read together, most of the savings come from orchestration, and the model adds a smaller amount on top.

Independent work points the same way. A September arXiv paper on enterprise coding agents argues that the harness decides which price applies and how many tokens are spent at that price.13 It finds that routing by list price gets long sessions wrong, because cache-read pricing dominates the cost of long agentic work.13 One practitioner write-up reports cutting compliance-alert processing from $2.89 to $0.25 per alert by moving repetitive steps from model calls into code.12 Another analysis found that the same model and same results can cost twice as much depending on the app wrapped around it. Its main example was first-call context of about 2,000 tokens in one harness versus about 27,000 in Claude Code.14 Writer is staking its business on this argument.

Governance and the trust question

The third part of the release puts a price on each automation. Administrators get a central view of agent activity, analytics for each of Writer's shareable Playbooks and Skills, and consumption alerts and spending limits that can be set by team.13 This completes a set of governance features Writer has been adding since last November.9

CEO May Habib's comments point to a sharper angle. She told TechCrunch that enterprise buyers are tired of chasing benchmarks and want stable costs. She also said rising token bills are making IT leaders distrust the big AI labs, which earn more when customers use more tokens.45 Writer stays model-agnostic: X6 sits alongside outside models brought in through Azure or Amazon Bedrock.4 That lets Writer present itself as the cost-control layer between enterprises and the frontier labs.

The takeaway

Writer's launch reflects a broader shift in enterprise AI. Frontier model quality is becoming easy to get through open weights, and Writer's view is that companies will compete on how well they orchestrate, govern and control costs. Writer's own numbers back that view: most of its savings come from the harness, not the model. The open questions are whether Writer's internal benchmarks hold up under independent testing, and whether enterprise buyers accept a Chinese-origin base model once Writer has post-trained it on U.S. infrastructure.