Tech • AI • Robotics • Game
Anthropic finally reveals how to prompt Fable 5.2!
7/10TL;DR
Anthropic’s guidance on Claude model selection argues that most teams can cut costs sharply by defaulting to Opus 5.5 at medium effort, reserving Fable 5.2 for the hardest long-running agent tasks and using cheaper models for narrowly defined subtasks.
KEY POINTS
Start with Opus 5.5
Anthropic recommends beginning with Opus 5.5 at medium effort, the default setting in Claude Code. That setup is presented as a strong everyday choice because it delivers reasoning close to Fable 5.2 while costing about half as much per token. For many coding and problem-solving workflows, that makes Opus 5.5 the baseline rather than the premium fallback.
Use Sonnet and Haiku for constrained work
Sonnet 5.5 is better suited to clearly specified jobs such as bug fixes, feature work from precise requirements, and tasks where the desired outcome is already mapped out. For high-volume operations like classification, routing and data extraction, faster low-cost models such as Haiku 1.5 are more appropriate. The broader message is that model choice should follow task structure, not habit.
Fable 5.2 is powerful but operationally harder
Fable 5.2 is positioned for the most complex projects, especially long-running autonomous applications and difficult coordination problems. Its main weakness is that it can stop midway through a task, requiring tighter prompting and stricter workflow design to complete work reliably. In return, it offers a more advanced execution architecture that can outperform Opus 5.5 on genuinely complex multi-agent jobs.
Prompting must be shorter and more structured
Effective use of Fable 5.2 depends on rewriting the system prompt, often stored in CLAUDE.md, so it is concise, explicit and free of personal background or stylistic clutter. The model benefits from a clear map of project directories, allowed dependencies and behavioral constraints. Guidance favors compact operational instructions over long identity-based prompts that describe preferences, biography or tone.
Architecture matters more than slogans
The advice rejects the common habit of telling a model to be an “expert” and handle everything alone. AI systems are described as probabilistic tools that need a defined decision framework, verification criteria and bounded responsibilities. Human operators still need to decide how memory is organized, how agents are scoped and how outputs are checked before action is taken.
Effort tuning can save more than model switching
The effort setting acts as a cost-performance dial. On SWE-bench Pro, Opus 5.5 at medium effort reportedly scores only 2.5 points below high effort while costing about 70% as much. Moving to very high effort adds just 1.4 points, but raises cost to roughly 2.5 times the high setting, suggesting that beyond a certain point it is cheaper to switch models than to keep increasing effort.
Sub-agents should use cheaper models
Multi-agent systems can reduce spending by assigning low-cost models to sub-agents with narrow, verifiable roles, such as reading logs, reviewing tests and reporting anomalies. More expensive models should remain in the coordinator role for code changes and decisions requiring stronger judgment. This approach makes pricing depend on the actual complexity of each subtask instead of charging premium rates across the full workflow.
Token and cache management are now core skills
Context delivery, prompt compression and cache behavior directly shape operating cost. Because cache writes can require paying to load an entire conversation context when switching to a new model, teams are encouraged to compact context aggressively and start fresh sessions with shorter plans when possible. Workflow design is becoming as important as raw model capability.
Autonomous AI remains expensive at scale
The economics of large agent systems remain steep. Figures cited for autonomous runs include about $428 for 16 agents over 8 hours, roughly $5,000 for 64 agents over 24 hours, and close to $200,000 for 1,000 agents over 72 hours. Those numbers underline why optimization, verification loops and careful model allocation are now central to enterprise adoption.
Labor market pressure is intensifying
The broader argument is that AI skill is no longer optional for knowledge work. A projected 30% decline in hiring of new graduates by 2026 is cited as evidence that employers will increasingly favor workers who can improve output while controlling AI costs. In that environment, value will come less from access to tools than from the ability to design efficient, reliable workflows around them.
CONCLUSION
The current generation of Claude models rewards disciplined orchestration more than brute-force use of the most powerful system. For companies and individual developers alike, the key advantage lies in matching model, effort level and agent design to the task while keeping token costs under control.
Ask a question