Orchestrating ChatGPT and Gemini from Claude Fable 5

I run Claude Fable 5 as the orchestrator on three projects now. In every one, its own coding output ends up being maybe sixty percent of the final diff. The rest comes from GPT-5.6 and Gemini 3.1 Pro, called as tools, not opened as chat tabs.

Forget multi-agent theater. This is one model with a system prompt that says you're the lead, here are two tools named ask_chatgpt and ask_gemini, use them when a second opinion beats a guess. It's wired through function calling, the same way you'd wire a search tool.

Why bother with three vendors

Single-vendor blind spots are the whole reason. Claude, GPT and Gemini are trained on overlapping but not identical data, tuned by different RLHF teams chasing different eval suites, so their mistakes don't correlate the way three runs of the same model would.

On a Stripe webhook integration last month, Fable 5 wrote the handler, Gemini flagged that the idempotency key wasn't scoped to the event type, and GPT-5.6 caught a currency-rounding edge case neither of the others mentioned. None of that required a human spotting three separate blind spots by hand.

The actual wiring

• Fable 5 as orchestrator, using Claude's native tool-use: each tool is a thin wrapper that calls the OpenAI Responses API or the Gemini API and returns plain text
• One tool per vendor, not one tool per task, keeps Fable 5's tool-selection reliable instead of second-guessing which of ten narrow tools to reach for
• A hard turn limit (I cap it at three round trips), or Fable 5 will happily ping-pong between models chasing marginal agreement that never arrives
• Normalize every response to plain markdown before it goes back into Fable 5's context, since each vendor's JSON schema and citation format differs enough to confuse the orchestrator otherwise

The orchestrator just holds the pen. It's no smarter than the tools it calls.

Where I actually use it

Cross-vendor orchestration earns its cost on high-stakes review, not on day-to-day generation. I reserve it for pre-merge review on anything touching auth, money or data migrations, the same instinct behind agentic architectures that don't fall over. For routine feature work I let Fable 5 run solo, since the extra API calls and latency aren't worth it for a CRUD endpoint.

Cost is the honest tradeoff here. A three-model review round trip runs three to four times the token cost of a single Fable 5 pass, plus the latency of two extra network hops if you run them serially. I run the two tool calls in parallel and only eat the cost on the review pass, never on drafting.

What actually breaks

The real limitation: none of the three models share memory. Every tool call carries the full relevant context again, so your token bill scales with how much you paste into each sub-call, not with how smart the orchestrator is. That adds up fast on a large file. See the architecture that cuts 99% of your LLM bill for the caching tricks that offset it.

The other failure mode is silent disagreement: when GPT-5.6 and Gemini both approve something Fable 5 got wrong, the orchestrator has no reason to doubt itself. Ensemble review catches uncorrelated mistakes, not the ones baked into every training set at once.