Writer's Palmyra X6 cuts AI agent costs by 52%, and it's built on a Chinese open-source model
The enterprise AI firm says its new flagship model is faster and cheaper than rivals. But the base it was built on is raising eyebrows.
Key points
- Writer launched Palmyra X6 on 14 July 2025, claiming a 52% average cost reduction and 48% speed improvement for enterprise AI agents.
- Palmyra X6 is post-trained from GLM-5.2, an open-weight model built by Beijing-based Z.ai, and runs entirely on US infrastructure according to Writer.
- The model has 744 billion parameters but was fine-tuned using just 626 curated training examples, all of them synthetically generated.
- Writer prices Palmyra X6 at $2 per million input tokens and $8 per million output tokens, compared to $15 and $75 for Anthropic's Claude Opus 4.8.
- Goldman Sachs forecasts global token consumption will reach 120 quadrillion tokens per month by 2030, driven almost entirely by always-on AI agents.
Enterprise AI platform Writer has launched Palmyra X6, its new flagship AI model, alongside rebuilt tools for managing and controlling AI agents. The headline claim: companies using Writer's agent platform can now expect costs to drop by 52% on average, tasks to complete 48% faster, and output quality to tick up by 10%.
Those are big numbers. But the story behind the model may matter more to enterprise buyers than the benchmarks.
What is Palmyra X6, exactly?
Palmyra X6 is not a model Writer built from zero. It starts from GLM-5.2, an open-weight model (meaning its internal settings are freely available to download and modify) released in June 2025 by Z.ai, a Beijing-based AI lab formerly known as Zhipu AI. Writer then applied its own training process on top, using a technique it calls anchored supervised fine-tuning, or ASFT.
The twist is how little extra training it took. Writer used just 626 examples, all of them machine-generated rather than written by humans, and trained the model for a single pass through that data. The goal was to teach the model new skills, like using external tools and delegating tasks, without accidentally eroding the general knowledge it already had.
"It's very much a Palmyra model, and we just happen to grab the floating point numbers as the starting point, and train from there," said Dan Bikel, who leads AI research at Writer.
The finished model has 744 billion parameters (think of these as the internal dials that shape how the model thinks) but only activates around 40 billion of them for any given piece of text. That architecture keeps it fast and relatively cheap to run.
Should enterprise buyers worry about the Chinese base model?
Freely, yes: this is a live debate in the industry. Using a Chinese-origin model as a foundation would have been a non-starter for US enterprise vendors just two years ago.
Writer is upfront about the provenance and insists the security concerns don't apply here. "This model is in no way, shape, or form connected to any of its original developers. It is fully run on our US infrastructure," Matan-Paul Shetrit, Writer's director of product management, told VentureBeat exclusively ahead of the launch. GLM-5.2 is released under the MIT licence, one of the most permissive open-source licences available, meaning anyone can legally download, modify, and redistribute it.
Whether that reassures a bank's compliance team is a different question, and one every enterprise buyer will need to answer for themselves.
How does the pricing stack up?
| Model | Input (per million tokens) | Output (per million tokens) |
|---|---|---|
| Palmyra X6 (Writer) | $2 | $8 |
| Claude Opus 4.8 (Anthropic) | $15 | $75 |
| Claude Sonnet 4.6 (Anthropic) | lower than Opus | lower than Opus |
| GPT-5.5 (OpenAI) | not disclosed | not disclosed |
On Writer's own internal benchmarks across nine task types, X6 scored 0.87 out of 1.00. Claude Opus 4.8 scored 0.86 and GPT-5.5 scored 0.80. Writer is candid that grading your own homework invites scepticism, and says its technical report covers the full evaluation methodology.
Why do AI agent costs spiral so fast?
A regular chatbot generates one answer per question. An AI agent, software that can carry out multi-step tasks on its own, breaks a single request into dozens of planning steps, web lookups, tool calls, and error checks, each one consuming metered tokens (the units of text an AI model reads and writes). The user sees one result; the bill reflects every step behind the scenes.
Goldman Sachs forecasts that monthly global token consumption will hit 120 quadrillion tokens by 2030, a 24-times increase from 2026 levels. Critically, falling per-token prices don't guarantee falling bills: if an agent uses 20 times more tokens while the unit price drops 75%, the total charge still rises fivefold.
Writer's pitch is simple: make each agent loop cheaper and faster, so companies can automate workflows that previously made no financial sense. "The biggest barrier today to enterprise expansion using AI is actually not model capabilities in most cases; it's actually the cost around them," Shetrit said.
Writer is used by Fortune 500 companies including Accenture, Uber, and Vanguard.
Common questions
Is my company's data safe if the underlying model comes from a Chinese lab?
Writer says Palmyra X6 runs entirely on US servers and has no ongoing connection to Z.ai, the original developers. The base model is open-source under the MIT licence. That said, your own legal and compliance teams should review the arrangement before deploying it for sensitive workloads.
How long can Palmyra X6 work on a task without human input?
Writer says the model can work unattended toward a single goal for up to eight hours, completing tasks in 26 seconds on average. That makes it suited to overnight batch processes and complex multi-step workflows.
Can smaller companies use this, or is it only for big enterprises?
Writer is an enterprise platform, so pricing and access are geared toward business customers rather than individual consumers. The token prices listed above apply when calling the model via Writer's API.
