Advisor Mode

A productized planner/executor split where the set model executes and a smarter advisor model is consulted only when the executor gets stuck. The executor (the model you have selected) does the bulk work — reads, writes, runs tools — cheaply; but any time it hits a wall it shares its context with the advisor and asks "here's where I am, I'm stuck, what should I do?", and the smarter model steers it back on course. This is the escalation-based, on-demand form of Model-Tier Routing: instead of the smart model dispatching sub-work up front, the cheap model runs continuously and pulls the expensive model's intelligence only at the moments it's needed — so you pay for premium reasoning at the decision points, not for every token of execution.

The source ("Chase AI", reducing Fable 5 usage) attributes advisor mode to an Anthropic "advisor strategy" that first shipped with Opus advisor + Sonnet executor: on the SWE-bench Multilingual chart the source shows, Sonnet 4.6 High + Opus advisor scores 74.8% at $0.96/task versus Sonnet 4.6 High solo at 72.1% at $1.09 — better and cheaper. The operational gotcha: whatever model you have set is the executor (it writes the code), so to make the top model the advisor you must set your model to the cheaper executor first (e.g. Opus) and then /advisor <model> (e.g. /advisor fable). The source notes Anthropic hasn't published Fable-as-advisor numbers, so the Fable-advisor case is presented as an extrapolation from the Opus/Sonnet result, not a measured one.

Claims

Advisor mode splits roles by escalation: the set model executes cheaply and consults a smarter advisor model only when it gets stuck — sharing context and asking what to do. principle — durable: it buys premium reasoning at the decision points rather than across all execution, which is the cost-efficient shape of a planner/executor split. It is Model-Tier Routing pulled on demand rather than pushed up front.
Whatever model is set is the executor (it writes the code); to make the top model the advisor, set your model to the cheaper executor, then /advisor <top-model>. best practice — context: this specific harness's advisor UX (the source demonstrates Claude Code); the exact commands are tool- and version-specific and will drift, but the role assignment (set-model = executor) is the load-bearing part.
A cheap executor with a smart advisor can beat the cheap model solo on both score and cost. observation — the source's SWE-bench Multilingual figures (Sonnet 4.6 High + Opus advisor 74.8% @ $0.96 vs Sonnet solo 72.1% @ $1.09) are reproduced as presented from Anthropic's "advisor strategy" chart; groundable, flagged for a later verification pass, not adjudicated here.
Applying advisor mode with the newest model as advisor is an extrapolation, not a measured result, when the vendor hasn't published numbers for that pairing. observation — the source is explicit that Fable-as-advisor lacks official numbers and reasons "by assumption" from the Opus/Sonnet graph; worth flagging so an inferred benefit isn't read as a benchmarked one.
The same source later argues against advisor mode: the executor slot should go cross-vendor rather than to a smaller model on the same ladder. observation — an unresolved tension, and one to state precisely because the two positions are narrower than they sound. Ten days after the distillate that put this page in the vault, Chase AI says: "why have Opus execute if at the same price I could have 5.6 do it?… I think this is way better than passing things off to Opus or to Sonnet or using advisor mode inside of Claude Code, because these GPT models are just better than those smaller Anthropic models and they are cheaper." What is not disputed: advisor mode's measured Opus/Sonnet result, or the planner/executor shape itself. What is disputed: the executor selection — the claim is that the in-vendor ladder is dominated by a cross-vendor option. Two conditions bound it: it requires holding both vendors' subscriptions (a Claude-only operator has no such option, and advisor mode remains live for them), and it rests on vendor-published benchmark claims that are unverified. Not adjudicated here; note also that this is one practitioner revising his own recommendation, not two independent sources disagreeing. See Cross-Model Independence for the other argument for going cross-vendor — one that does not depend on the rival model being better at all.

Model-Tier Routing — the parent pattern; advisor mode is its escalation-based (pull-when-stuck) instance, versus the planner dispatching sub-work up front.
Reasoning Effort Control — advisor mode spends premium capability at decision points; effort control spends it via the reasoning dial. Both target "pay for intelligence only where it earns its keep."
Agent Supervision — structurally the mirror image: supervision is a human advising an agent when it needs steering; advisor mode is a smarter model playing that role for a cheaper one.
Agent Loop — the executor's read/write/run loop with an escalation hook to the advisor when stuck.
Cross-Model Independence — the structural objection to this whole shape: advisor and executor drawn from one vendor's ladder share a lineage, so escalation buys capability but not independence — the advisor's blind spots are the executor's blind spots.
Distillate: Make Fable 5 80% Cheaper (& Other Usage Cheat Codes) — five levers for spending a premium model less
Distillate: Fable 5 + GPT 5.6 Sol = CHEAT CODE — the same channel arguing the executor slot should be filled cross-vendor instead; a tension to resolve, scoped to executor selection, not to the pattern.

Advisor Mode

Claims

Related

Linked from