Writer Says Palmyra X6 Cuts AI Agent Costs 52%
Writer says Palmyra X6 cuts AI agent costs 52%, using GLM-5.2 open weights, a faster orchestration harness and new enterprise spending controls for CIOs.
Summary
Writer released flagship Palmyra X6, a rebuilt Writer Agent harness and governance controls on August 13, 2026 for customers including Accenture, Uber and Vanguard. It claims 52% lower cost, 48% faster work and 10% better quality. X6 charges $2 per million input tokens and $8 per million output tokens, versus Anthropic Claude Opus 4.8's $15 and $75, averages 26 seconds per task and runs unattended for eight hours. Writer's nine-capability test scored X6 0.87, Opus 4.8 0.86, Claude Sonnet 4.6 0.85, OpenAI GPT-5.5 0.80 and Google Gemini 3.1 0.77; methodology appears in its technical report.
The 744-billion-parameter mixture-of-experts model activates roughly 40 billion per token and retains GLM-5.2's architecture. Writer post-trained the June, MIT-licensed open-weight release from Beijing's Z.ai, formerly Zhipu AI, on U.S. infrastructure using anchored supervised fine-tuning: 626 filtered synthetic trajectories, one low-learning-rate epoch, token weighting, a KL-divergence anchor and Muon optimization. Its pre-registered risk study covered 19,674 blinded-judged responses; X6 showed both sides on ModelSlant 80% of the time and its system message lifted FORTRESS safety 8.6 points over raw GLM-5.2 with negligible helpfulness loss. English political asymmetry lacked statistical robustness, but behavior varied by language. SaferAI found base GLM-5.2 refused none of its offensive cyber or biology tests, while Z.ai published no safety framework or pre-deployment assessment.
The harness independently cut cost 41% and task time 44% without quality loss across Writer, Anthropic and OpenAI models. Writer Agent added Anthropic, OpenAI, Microsoft Azure, AWS Bedrock, Nvidia NIM and image-model support; dashboards provide companywide and per-Playbook or Skill usage, alerts and limits. Goldman Sachs forecasts monthly token consumption will rise 24-fold from 2026 to 2030 to 120 quadrillion; 20 times more tokens with prices down 75% still makes a task five times costlier. Writer was founded in 2020 by May Habib and Waseem AlShikh and raised $200 million at a $1.9 billion valuation in late 2024.
Positives
- 52% lower average cost, 48% faster execution and 10% higher quality are Writer's headline gains for Writer Agent paired with Palmyra X6.
- $2 input and $8 output rates per million tokens sharply undercut Claude Opus 4.8's respective $15 and $75 prices.
- 0.87 on Writer's nine-capability evaluation placed X6 narrowly ahead of Claude Opus 4.8 at 0.86.
- 41% lower cost and 44% faster completion came from the rebuilt harness alone across Writer, Anthropic and OpenAI models.
- 19,674 blinded-judged responses supported Writer's risk evaluation of political bias, censorship, factuality and refusal behavior.
- Multi-model support gives administrators access to Anthropic, OpenAI, Microsoft Azure, AWS Bedrock, Nvidia NIM and image models.
Risks & concerns
- 120 quadrillion monthly tokens by 2030 would represent a 24-fold increase from 2026, according to Goldman Sachs.
- 20 times greater token use still raises task charges fivefold when token prices fall 75%.
- SaferAI found GLM-5.2 refused none of its tested offensive cyber or biology requests through Z.ai's public API.
- Z.ai published no safety framework or pre-deployment risk assessment for the GLM-5.2 base model, according to SaferAI.
- Language-dependent political behavior remained after Writer's 626-trajectory post-training process, despite no statistically robust English asymmetry.
- Writer's leading quality scores come from internal evaluations, creating unavoidable skepticism despite publication of its methodology.