Detailed Analysis
A Reddit user has surfaced a concerning bug report involving what appears to be Anthropic's orchestration and model-selection interface, describing a scenario where sessions labeled as "Sonnet 5 - Max" in the UI were, in actuality, running on a different model tier referred to as "Fable 5." The user reports that an orchestration session—one designed to spawn subsidiary sessions to implement features—defaulted new spawned sessions to Fable 5 regardless of the model explicitly selected by the user. Critically, even after manually switching the displayed model to Sonnet 5, the backend continued executing on Fable 5, while the UI's bottom-right model indicator falsely displayed Sonnet 5. This mismatch between displayed state and actual execution state caused the user to unknowingly burn through their Fable 5 usage quota on routine "grunt work" tasks that were never intended to consume that particular allocation.
This type of bug strikes at a fundamental trust issue in AI tooling: the reliability of state indicators within a user interface. When users select a specific model for a task—often for reasons of cost, capability, or quota management—they are relying on the UI to accurately reflect what is actually running "under the hood." A silent mismatch between the selected/displayed model and the executed model undermines a core promise of these tools: that users have granular control over resource allocation and task routing. In multi-model ecosystems where different tiers (e.g., a faster/cheaper model versus a more capable "Max" one) carry different quota costs, this kind of bug isn't merely cosmetic—it has real financial and productivity consequences, as evidenced by the poster's frustration at losing precious quota to work that didn't need premium-tier reasoning.
The mention of "orchestration" sessions capable of spawning subsidiary sessions to implement features reflects a broader trend in agentic AI coding tools: multi-session, multi-agent architectures where a top-level session delegates work to child sessions, potentially running on different models or configurations optimized for different subtasks. As these architectures grow more complex, ensuring consistent and accurate state propagation—especially around model selection, which directly ties to quota and billing—becomes increasingly difficult but also increasingly critical. Bugs like this one suggest that as orchestration features mature, the interplay between session inheritance, model overrides, and UI rendering needs more rigorous testing, particularly around edge cases where a user attempts to override a default after a session has already spawned.
More broadly, this incident is emblematic of growing pains in the rapidly evolving space of AI coding assistants and agent frameworks, where new features (like automatic orchestration and spawning) are shipped quickly to keep pace with competitive pressure, sometimes outpacing the polish of the underlying UI/state-management layer. For users, it serves as a practical warning to manually verify actual model usage (through logs or usage dashboards, if available) rather than trusting on-screen indicators alone, especially when working with quota-limited premium models. For Anthropic and similar providers, reports like this represent valuable, if painful, bug intelligence that underscores the importance of transparent, auditable model-routing telemetry as agentic multi-session architectures become a standard feature of AI development tools.
Read original article →