Fable 5 as Advisor: Anthropic's Two-Model Pattern for Smarter, Cheaper Agents (Ep. 1010)
The episode explores Anthropic's "advisor strategy," a novel AI agent pattern that combines a fast, cheap "executor" model with a frontier-class "advisor" model. This allows for mid-task consultation within a single API call, resolving the tension between cost and capability. Benchmarks show simultaneous improvements in quality and cost reduction, demonstrating that AI progress is shifting towards intelligent model composition rather than just larger models.
This episode delves into the “advisor strategy,” a novel pattern for building more cost-effective and faster AI agents. It addresses a fundamental tension faced by AI agent developers: frontier models offer superior planning and error recovery but are expensive, while smaller, faster models are cheaper but often make critical errors in decision-making.
Anthropic’s Advisor Tool
Anthropic has productized this strategy with their advisor tool, available in beta via the Claude API. This tool allows for pairing two models within a single API call:
- The Executor: A fast, cost-efficient model (e.g., Claude Sonnet, Haiku) that drives the agent loop, reading tool results, deciding next actions, and generating final output.
- The Advisor: A higher-intelligence model (e.g., Claude Opus, Fable) that the executor can consult mid-task, much like calling any other tool. The analogy is a lower-level worker seeking guidance from a supervisor.
Mechanically, the advisor is declared as a tool. When the executor needs help (e.g., before committing to a design or after repeated errors), it emits an advisor call without parameters. Anthropic’s servers automatically provide the advisor with the full conversation transcript, including the system prompt, tool definitions, all prior turns, tool results, and the executor’s current generation. The advisor runs a separate server-side inference pass, generates typically 400-700 tokens of advice (with internal thinking stripped out), and the executor then continues. Crucially, this entire process occurs within one API request, eliminating the need for external orchestration frameworks, extra network round trips, or context switching overhead, as both models share the same conversation state.
Performance Benchmarks
Anthropic’s benchmarks highlight significant improvements:
- Sonnet with Claude Opus Advisor: On SweBench multilingual (software bug fixing), Sonnet with an Opus advisor scored 2.7 percentage points higher than Sonnet alone, while simultaneously reducing cost per agentic task by 12%. This demonstrates improved quality at a lower cost.
- Haiku with Claude Opus Advisor: On BrowseComp (web research), Haiku alone scored ~20%. With an Opus advisor, its score more than doubled to 41%. While still trailing solo Sonnet in quality, this configuration cost 85% less per task than Sonnet alone. The pattern is clear: the weaker the baseline executor, the greater the lift from an advisor.
- Sonnet 5 with Fable 5 Advisor: On SweBench Pro (482 tasks), Sonnet 5 alone passed 75% of tasks at $0.75/task. Adding Fable 5 as an advisor (consulted ~once per task) boosted the pass rate to 84% at $1.40/task. Fable 5 solo achieved 92% at over $2/task. The Sonnet+Fable pairing achieved 92% of Fable’s standalone performance at 63% of its cost.
Practical Considerations
- Use Cases: The advisor strategy is not suitable for single-turn question answering. It excels in long-horizon agentic tasks like coding, computer use, and multi-step research, where planning is crucial.
- Executor Behavior: Executors tend to under-call the advisor. Anthropic suggests prompt language to encourage calls at critical moments: early (before committing to an approach) and late (before declaring task completion). Nudges significantly improved Haiku’s pass rates (7 percentage points) but had little effect on Sonnet and slightly lowered Opus’s, which are better at knowing when to seek help.
- Cost Containment: Setting a token cap (e.g., 2048 tokens) on the advisor’s output reduced mean output by roughly sevenfold without detectable quality loss. Advisor-side prompt caching is beneficial for long agent loops (breaks even around three calls per conversation).
- Model Capability: The advisor must be at least as capable as the executor.
Comparison with OpenAI’s Router
This strategy contrasts with OpenAI’s router, which makes a dispatch decision before generation based on query complexity and user intent. OpenAI’s approach is platform-driven for consumer chat. Anthropic’s advisor, conversely, allows the cheap executor model to run the show and escalate itself mid-task based on its own judgment, leveraging the full shared context of prior attempts.
Broader Implications
- The advisor pattern is spreading beyond Anthropic; OpenRouter now offers a cross-provider advisor tool, allowing any model to be an executor and any other model from any provider to be an advisor, leveraging significant price differences between model tiers.
- This approach productizes what many teams have been hand-rolling as escalation logic.
- It represents one of the most cost-effective experiments in AI engineering with potentially large payoffs.
- The deeper significance is that frontier AI progress is increasingly about smarter economics and composing existing models, rather than solely training bigger ones. It mimics how well-run human teams operate, making complex judgment available as a simple API parameter.