Writer Launches Palmyra X6 to Cut Enterprise AI Token Costs

Enterprise AI company WRITER is making a direct play for one of the biggest problems emerging as businesses move AI agents from experiments into everyday operations: the cost of keeping those agents running.

The company has introduced Palmyra X6, its new flagship agentic model, alongside a major upgrade to the orchestration system behind WRITER Agent. Rather than focusing only on making the underlying model more capable, WRITER says the combination is designed to reduce how many tokens an agent consumes while completing real business tasks.

That distinction is becoming increasingly important. A chatbot may generate a single response, but an AI agent can plan a task, search company data, call multiple tools, retry failed actions and delegate work to other agents. Each additional step can increase token consumption and ultimately the cost of completing the job.

Palmyra X6 is built around cost per task

Palmyra X6 was post-trained on top of Z.ai's GLM-5.2 open-weight model and optimized for enterprise workflows such as research, marketing content, revenue operations, retrieval, tool use and multi-step automation. WRITER's own evaluations cover nine capabilities including grounding, retrieval, tool use, content generation, sub-agent delegation and brand voice.

According to WRITER, X6 achieved an average capability score of 0.87 out of 1.00 in its internal evaluations. The company compared that with Claude Opus 4.8 at 0.86, Claude Sonnet 4.6 at 0.85, GPT-5.5 at 0.80 and Gemini 3.1 at 0.77. These are WRITER's own production-oriented evaluations rather than independent industry benchmarks, so the comparisons should be viewed in that context.

The pricing difference is notable. Palmyra X6 costs $2 per million input tokens and $8 per million output tokens. WRITER's comparison lists Claude Opus 4.8 at $15 and $75 respectively, while Claude Sonnet 4.6 was measured at $3 and $15.

X6 also comes with a 1 million-token context window, and WRITER says it can stay focused on a single objective for as long as eight hours without supervision. On the company's internal set of 22 production tasks, median cost per completed task was $0.12, while median task latency was about 26 seconds.

The bigger change may be the harness around the model

WRITER's argument is that companies may be focusing too heavily on which model they choose while overlooking the software that determines how that model actually works.

Its upgraded Agent harness decides how much context to send, when to use tools, when to delegate work and whether a request really needs extended reasoning. For simpler jobs, the agent can answer directly instead of running an unnecessarily long reasoning process. More complex work can be executed in parallel, divided among sub-agents or processed in batches.

WRITER published research in July testing that idea across 22 tasks and six foundation models. Keeping the models constant while replacing a conventional orchestration loop with WRITER's harness reduced average tokens per task from 14,200 to 8,800, a 38% decline.

Average cost per task fell 41%, from $0.21 to $0.12, while median completion time dropped 44%, from 48 seconds to 27 seconds. Task-completion scores remained broadly similar, moving from 0.78 to 0.81. The study also reported an 82% improvement in quality per dollar, while completed tasks per million tokens rose from 54.9 to 92.0.

When the upgraded harness is paired specifically with Palmyra X6, WRITER claims an average 52% reduction in cost, 48% improvement in speed and 10% improvement in quality compared with its earlier configuration.

Enterprise AI is moving beyond the benchmark race

The release points to a broader shift in enterprise AI. Companies increasingly have access to many highly capable models; the harder problem is deciding how to deploy them economically across thousands or millions of automated tasks.

WRITER is also keeping its platform multi-model. Administrators can enable models from providers such as OpenAI and Anthropic, while businesses can bring models through services including Microsoft Azure, AWS Bedrock and NVIDIA NIM. The company has also added consumption analytics, token alerts and spending limits so IT teams can see which agent workflows are actually creating the biggest bills.

That may ultimately be the more consequential part of the Palmyra X6 launch. The next stage of enterprise AI competition may not simply be about producing a smarter model. It could increasingly be about how much useful work a system can finish for every dollar and every million tokens consumed.

If WRITER's efficiency results hold up across larger independent deployments, Palmyra X6 and its upgraded harness offer a glimpse of where the agent market is heading: away from unlimited token consumption and toward AI systems whose economics can survive deployment at enterprise scale.