Writer prices Palmyra X6 against the invoice, not the leaderboard

Writer released Palmyra X6 on Thursday alongside upgrades to its agent platform, and the headline figures are financial rather than qualitative: the company says the platform now runs at 52% lower average cost with a 48% speed improvement and a 10% quality improvement when paired with the new model ^[1][2]. For anyone already running agents in production, that is the relevant axis, because the constraint on agent deployment stopped being capability some time ago and became the monthly bill.

Writer's own framing is explicit. "The enterprise wants token consumption to explode -- it means adoption is happening -- but they need costs to flatten," co-founder and CTO Waseem AlShikh said ^[3].

The pricing supports the pitch more clearly than the eval scores do. X6 averaged 0.87 across nine evaluations at $2 per million input tokens and $8 per million output tokens, against Claude Opus 4.8 at 0.86 for $3/$15, GPT-5.5 at 0.80 for $5/$15, and Gemini 3.1 at 0.77 for $2.50/$10 ^[4]. The quality gap over Opus is 0.01 ^[5], which is noise for most workloads. The price gap is not: X6 is 47% cheaper on output tokens and 33% cheaper on input ^[6], which works out to roughly 1.9 times the eval score per output dollar ^[7]. GPT-5.5, on these figures, costs 2.5 times as much on input and 1.9 times as much on output while scoring lower ^[8]. Gemini 3.1 is priced 25% above X6 on both sides of the meter ^[9].

The throughput numbers are where operators should do arithmetic. Writer says X6 completes tasks in 26 seconds on average, generates 82 tokens per second, and can work unattended toward a long-term goal for up to eight hours ^[10]. An average task therefore emits on the order of 2,100 output tokens ^[11]. An eight-hour run generating continuously would emit about 2.36 million output tokens, or roughly $19 at list price ^[12] -- an upper bound rather than a forecast, since agents spend much of a long run waiting on tools, but it is the right order of magnitude to budget against per long-horizon job.

Two caveats. First, the model is post-trained on top of GLM 5.2, the open-source mixture-of-experts model from Z.ai ^[13], so the cost structure rests on someone else's pretraining run as much as on Writer's engineering. Second, the 52%, 48% and 10% figures are Writer's own and the report does not state what baseline, workload mix or time period they are measured against ^[14], nor does it name the nine evaluations ^[15]. Treat them as vendor arithmetic until reproduced.

The platform changes are the more durable part of the release. Writer added reporting and governance controls giving administrators a view of activity across teams, plus shareable Playbooks and Skills -- repeatable agent workflows surfaced in a dashboard ^[16]. Admins get a dedicated analytics and governance section covering usage, ownership, and per-task or per-skill cost, and can set consumption thresholds and alert limits ^[17]. That is chargeback plumbing, and it is what makes a 52% claim auditable inside a customer rather than just quotable.

Watch whether independent evaluations reproduce the 0.87, and whether the political-bias claim Writer makes for X6 ^[18] survives outside contact with its own test set.