Claude Code already lets my default model, Claude Opus 5.5, consult a bigger one: with /advisor set to Claude Fable 5.1, Fable returns guidance that Opus applies before continuing. Anthropic measured that pairing on an internal coding benchmark at 1.7 points over Opus 5.5 alone for about 2.1 times the money, “about what more effort buys.” For the work I escalate, it leaves four gaps:
| Gap | What the advisor does | Why it matters to me |
|---|---|---|
| Vendor | Every advisor it accepts is a Claude model. | A second vendor gives me a second, independent read. The advisor can only give me more Claude. |
| Timing | ”Claude decides when to call it.” I can ask for a consultation, but no setting forces one. | I’m the one who knows which decision is worth the extra cost. Anthropic’s own cost guide calls how often the model asks “the fragile one”. |
| Input | It “receives the full conversation, including every tool call and result.” | My framing goes with it. Whatever the session has already leaned toward, the advisor reads too. |
| Score | Nothing records whether the call paid. | Anthropic scored the pairing on coding tasks with right answers. My design and copy questions have none, so only my own picks can show whether escalating was worth it. |
I built a Claude Code skill called panel to close those four gaps. When a decision deserves more than one take, I ask my session to run a panel. It writes one brief and sends it to two bigger models from two vendors, GPT-6 Astra and Claude Fable 5.1, and to my default model as a blind control. A small script handles the parts that shouldn’t depend on how good the models are: the launch, the blind labels and the record of what I pick. This post was its first real run: I took sections from all three answers, and the opening came from the control.
For design questions, the session builds the strongest three or four options on one comparison page; for copy, it puts the strongest versions side by side.
Every route Anthropic offers stays inside Claude: the advisor, a subagent pinned to Fable, the escalation pattern in Managed Agents. I had already sent GPT-6 Astra work six times in eight days through OpenAI’s Codex CLI, with a new script each time. The panel sends one brief to Astra at its ultra reasoning setting and to Fable in the same launch. When two vendors name the same cause without seeing each other’s answer, that agreement is the strongest signal on the page. On the first design question I sent both, they did: the small uppercase date under each title in this site’s post list competed with it. The script also reads each vendor’s own log to confirm which model answered, because a mistyped Codex effort setting falls back to its default without a word.
I can ask Claude for an advisor consultation, but the docs say there is no setting to force one. The panel starts when I run it. I choose which decision deserves the extra consultation.
The panel gets one neutral brief my session writes instead of my conversation: my words, everything already tried with my verdict on each, the evidence, the constraints, and the axes the options can vary on. Each panelist sees only that folder, with read-only tools, no API keys and none of my connected services. Every answer comes back in one shape, a diagnosis, two or three options valued on each axis, a pick with its weakest point, and a list of the figures it estimated, under a random label, P1 to P3, so I can’t tell which is the big model until I’ve read them all.
The same brief also goes to my default model, Opus 5.5, as a blind control. My pick goes into a ledger, which counts across runs how often each panelist’s option gets chosen, the control’s included.
I suspect the panel will earn its cost on design and copy, where a spread of options matters, and lose to the control on plain decisions. The ledger tags every run by those types, so the next rows can prove me wrong.