David Ondrej says Fable 5 feels more capable in real coding projects because it follows the deeper intent of a task and tests its work more consistently. He reports that using it through Cursor produced relatively few downgrades after access returned, although safety classifiers still affect some cybersecurity-related language.
His recommended architecture uses Fable 5 as an orchestrator, planner or manager while smaller open models execute routine steps. That preserves the strongest model for decisions that need context and judgment, while models such as GLM, Kimi, MiniMax and DeepSeek handle cheaper implementation work.
Ondrej also recommends using a strong model to create reusable skills and long-term plans that less capable agents can follow later. He argues that products should expose clean command-line tools, APIs and documentation because agents will increasingly operate software on behalf of people rather than relying on human interface navigation.
The longer-term plan is resilience through open weights, retained datasets, provider diversity and eventually self-hosted compute. Ondrej views that capacity as protection from access changes and rising token costs, while still reserving frontier models for work where their additional capability produces real value. A distinct Accio Work sponsorship and promotional calls for courses and downloads are omitted from the catalogue record.
Watch on YouTube
