💡 Claude Fable 5 Deep Work Mode: Autonomous Multi-Hour Agent Sessions via Mythos-Class Model

[github-actions[bot]]

Summary

Introduce a "Deep Work" session type powered by Claude Fable 5, Anthropic's new Mythos-class model purpose-built for long-running autonomous agent tasks ($10/$50 per M tokens). Users describe a complex research or analysis task, the avatar estimates scope and cost, then runs autonomously for hours — checking in periodically via OS notifications and delivering a comprehensive result when complete.

Market Signal

Claude Fable 5 launched July 2026 as a fundamentally new model class designed for multi-hour autonomous operation. ServiceNow and NVIDIA's Project Arc (May 2026) demonstrated enterprise demand for self-evolving desktop agents that run for hours with Jensen Huang's keynote endorsement. Anthropic's Managed Agents API (/v1/agents, /v1/sessions, /v1/environments) provides cloud runtime infrastructure for stateful long-running sessions billed on session-hours. No consumer-facing app surfaces Fable 5's Mythos-class autonomous capabilities yet — this is a greenfield opportunity.

User Signal

Existing ideas #232 (Goal Mode with checkpoints) and #247 (Async Task Review Queue) express demand for background workflow execution. Idea #55 (Effort Modes) proposes model tier selection but predates Fable 5's release and its qualitatively different autonomy profile. The PRD's system notification strategy already envisions this pattern: notify on long-running task completion, prompt to continue unfinished workflows. The target persona (Sarah the PM) would benefit most: "Research our top 5 competitors and prepare a comparison report" as a fire-and-forget task that runs while she's in meetings.

Technical Opportunity

The AgentBackend abstraction layer supports model selection — adding Fable 5 is a configuration change. The Claude Agent SDK's session persistence enables resume across app restarts. The architecture's session management (SQLite storage, durable state on graceful/unexpected termination) handles the long-running lifecycle. Electron's native OS notification API delivers periodic check-in updates. The OutputPanel's TaskProgress display mode (FR43) provides real-time visual tracking when the user returns to the app. Fable 5 at $10/$50 per M tokens is positioned between Sonnet and Opus pricing, making multi-hour sessions economically viable for BYOK users.

Assessment

Dimension	Score	Rationale
Feasibility	high	Model config change + session lifecycle extension; architecture already supports the pattern
Impact	high	Unlocks qualitatively new workflows (multi-hour autonomous research/analysis) no competitor offers
Urgency	med	Fable 5 is available now but TalkTerm is pre-implementation; align with Phase 2 roadmap

Adversarial Review

Strongest objection: Multi-hour autonomous sessions at $10/$50 per M tokens could generate large, unexpected bills for BYOK users. A 4-hour deep research session might consume 500K+ tokens ($25+), which non-technical users may not anticipate.

Rebuttal: The avatar provides upfront cost estimation: "This will take about 2 hours and cost approximately $15-25. Want me to proceed?" The maxBudgetUsd SDK parameter provides hard guardrails. The confirm-plan pattern (FR20) gates session start. Existing cost meter ideas (#69, #411) ensure real-time visibility. Cost transparency builds trust by setting expectations — budget caps prevent runaway costs.

Suggested Next Step

Verify Claude Fable 5 API availability in the Claude Agent SDK. Design the "Deep Work" session type UX: task description → scope estimation → cost preview → approval → periodic check-ins via OS notification → result delivery with avatar narration. Define check-in cadence (e.g., every 30 minutes or at milestone completion). Integrate with existing session management and cost tracking infrastructure.