01
What Optimizely announced
Optimizely announced a new family of purpose-built, post-trained AI models for marketing on September 1, 2026. The company says its Agent Platform can route work to specialized models instead of relying on one general-purpose frontier model for every task.
Optimizely has not published individual model names, model IDs, parameter counts, architecture details, standalone pricing, or a public API reference for this model family. Its launch announcement says the models are coming soon to all Agent Platform customers, so they should be treated as announced rather than generally available.
02
What Mark-Bench measures
Alongside the model family, Optimizely introduced Mark-Bench, which it describes as an open-source benchmark for evaluating AI systems on marketing work. The company says it spans 285 tasks across 15 marketing functions and more than 6,000 evaluation criteria.
- Writing a press release
- Creating a social post
- Writing email copy
- Evaluating both cost and task performance
Optimizely says the benchmark is intended to let marketers, researchers, and AI providers compare models and agentic harnesses on domain-specific work. A verified public repository or install guide was not located during this publication check, so this article does not claim that readers can already clone or run Mark-Bench.
03
How to interpret Optimizely's benchmark claims
Optimizely reports that its purpose-built models achieved 10x greater cost efficiency than state-of-the-art LLMs in early testing. It also reports that, under Mark-Bench default configurations, Optimizely Agent Platform achieved a 67% all-pass rate versus 60% for Claude Code at 2x lower cost.
The narrower takeaway is that Optimizely is betting specialized post-training and task-level model routing can reduce unnecessary model size, context use, and token spend for recurring marketing jobs while maintaining useful quality.
04
How Mark-IQ and Agent Platform fit in
Optimizely says the new models draw on Mark-IQ, the Agent Platform's organizational context layer. That layer can incorporate information such as experimentation history and web analytics so agents work with brand and business context without rebuilding it in every prompt.
This model layer is separate from Optimizely Virtual Teammates. Virtual Teammates are persistent role-specific AI coworkers; the new model family is the specialized model layer intended to perform particular marketing tasks efficiently.
Optimizely's current Agent Platform documentation also describes a multi-model architecture that matches tasks to models and supports organizational context, agent workflows, connectors, and governance controls.
05
Pricing, availability, and important limitations
Optimizely has not announced standalone per-token or per-model pricing for this family. The 10x efficiency figure is therefore not a public price list and should not be converted into a customer price estimate.
Several important details remain undisclosed: individual model names, architecture, parameter counts, training data, licensing, API identifiers, standalone pricing, and an exact rollout date. Independent Mark-Bench results were also not identified during this review.
Until those details are published, the most accurate description is that Optimizely has announced a specialized marketing-model family and benchmark, with broader Agent Platform availability promised soon and performance claims still attributable to the vendor.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.