Writer rebuilds its agent harness around cheaper, longer unattended work

Writer released Palmyra X6 on Thursday, pairing its new flagship model with a rebuilt agent harness and pointing both at the problem enterprises complain about most: agent bills that grow faster than agent output. The company says its Agent platform, running on X6, now completes work at 52 percent lower cost, 48 percent faster, and with 10 percent better quality than the prior generation, SiliconANGLE reported. All three figures are Writer's own measurements.

The model is priced at $2 per million input tokens and $8 per million output tokens, and Writer cites an average task completion time of 26 seconds at 82 tokens per second. The more consequential claim for orchestration buyers is endurance: X6 can work unattended toward a long-horizon goal for up to eight hours, a full workday of autonomous execution without a human checkpoint. That places Writer in the long-running agent race that has defined this summer's platform releases, though with a mid-market price tag rather than a frontier one.

Notably, Writer did not pretrain from scratch. SiliconANGLE reports that X6 was post-trained on GLM 5.2, the open-source mixture-of-experts model from Z.ai, with Writer's value concentrated in the post-training, the harness, and the surrounding controls. Writer scored the result at 0.87 across its nine-evaluation internal suite, a hair above the frontier models it compared against, figures that, again, come from the vendor.

The harness work is where the cost claims live. Writer rebuilt how its agents plan, batch, and reuse context between steps, arguing that most enterprise agent spend is burned on redundant reasoning rather than useful work. The release also adds governance surfaces: centralized Playbooks for repeatable workflows, Skills dashboards, and administrator views of usage, performance, and spend, with consumption alerts across teams. VentureBeat, which covered the launch, framed the release around surging enterprise token consumption.

Co-founder and chief technology officer Waseem AlShikh told SiliconANGLE that enterprises actually want token consumption to grow, since it signals adoption, "but they need costs to flatten." That sentence is close to a thesis statement for the current phase of the agent market. The first wave of enterprise deployments proved agents could do the work. The renewal conversations now underway are about whether anyone can predict what the work will cost.

Writer's answer, cheaper tokens inside a stingier harness with spend controls bolted on, is not the only one. But the company is unusual in publishing the claim as a single number, and 52 percent is now a benchmark its competitors will be asked to beat.