Built for the long run.

What model powers WRITER?

Palmyra X6 — WRITER’s most capable model — is the default model across the WRITER platform. It’s built for the work marketing and revenue teams actually run: grounded in your company knowledge, able to hold your voice, able to call the tools your stack already runs on, and coherent across hours of work rather than minutes.

X6 shipped alongside a new version of WRITER Agent. The two were developed together — the model was built to run inside the agent, and the agent was built around the model.

Is Palmyra X6 the only model I can use?

No. X6 leads our evaluation, so it’s the default — but it isn’t the only option.

Admins can enable models from other providers, and users can choose a model at the start of a session. Specialized models can be selected for specific capabilities, such as image generation. And beyond WRITER’s own catalog, admins can bring their own models from AWS Bedrock, Microsoft Azure, and NVIDIA NIM.

How was Palmyra X6 evaluated?

Public benchmarks measure what’s easy to grade. They don’t measure whether a model can pull twenty-two sources through a knowledge graph and cite each one, hold a voice across a 3,000-word draft, or hand half a research task to a sub-agent and merge the results back correctly.

So we built the evaluation out of the work itself — production tasks our customers run across marketing, revenue, research, and outreach — graded 0 to 1 against a locked baseline across nine dimensions, each chosen because it’s a common failure point in enterprise deployments. X6 improved on all nine dimensions, with no regressions.

The scores measure Palmyra X6 running inside WRITER Agent — the model plus the system around it — because that’s how customers run it.

How does X6 compare to other frontier models?

We ran the same evaluation against five other frontier models. X6 recorded the highest capability score in the set at the lowest frontier price. Models priced two to nine times higher scored no better on this work, and the one materially cheaper model scored considerably worse.

What does it cost to run?

X6 costs $2 in / $8 out per million tokens. It finishes the average task for roughly $0.12 — about half the per-task cost of the previous generation, in about half the time. Output tokens cost 9.4x less than the most expensive frontier model in our comparison set.

Per-task cost is the number that matters, because the unit of work has changed. It used to be a prompt and a response; now it’s an objective handed to an agent that plans, researches across twenty sources, drafts, checks its own work, and returns something finished. A model priced for occasional high-stakes questions becomes a different proposition entirely for a marketing team running four hundred campaigns a quarter.

How long can an agent run on a single objective?

Running inside WRITER Agent, X6 holds one goal for up to eight hours — planning, executing, testing its own output, correcting, and delivering finished work rather than a draft someone has to supervise.

Most models drift off a long objective within minutes. The difficult part is holding intent, not just context, across hours rather than turns — and it’s what makes background automation practical: agents left running to watch for market signals or competitor moves and act when something happens. Because each step is efficient, those runs are inexpensive enough to leave on.

Do you train on our data?

No. Your data never trains our models. Prompts, completions, and documents stay yours, with zero retention by default. The playbooks, workflows, and institutional knowledge your teams build remain yours.

Guardrails screen output for PII and copyrighted material before it leaves the platform. Administration includes SSO/SCIM, granular RBAC, audit logs, and customer-managed encryption keys, aligned to the NIST AI RMF and the EU AI Act. WRITER holds SOC 2 Type II, ISO 27001, ISO 42001, HIPAA, GDPR, and PCI-DSS. Self-hosting and full data residency are available.

How do you handle political and ideological slant?

Marketing and revenue content goes out under your brand, to customers and prospects — so slant isn’t hypothetical, it’s a brand risk, and one that’s hard to catch by hand once you’re generating thousands of pieces a month.

We evaluated Palmyra X6 against seven other frontier models across eleven benchmarks, including the Washington Post’s ModelSlant eval. X6 presented both sides of a hot-button political question in 80% of its answers — the most even-handed result tested, well ahead of the next-best score of 57% — while also posting the field’s second-lowest refusal-mismatch rate at 1.2%. Full methodology and per-model results are in the transparency report.