Orchestration
Decomposing briefs, dispatching subagents, and owning the judgement calls at the top of the stack.
The one pattern every serious practitioner has independently converged on - we counted nine prominent creators saying it in their own words - is this: plan with the expensive model, execute with the cheap one. The orchestrator is the expensive seat. It decomposes the brief, dispatches the workers, and owns every judgement call the workers aren’t qualified to make.
Why Opus 5 takes gold
A close call, and the closest thing this board has to a coin flip - but on balance Opus wins for most scenarios, owing to Fable’s eye-watering cost. Opus tops the on-axis measures (Artificial Analysis Agentic Index #1; roughly 1.5× the next model’s pass rate on Zapier’s automation bench for the same money), posts a lower cost per task, and ships the thing an all-day seat actually needs: an effort dial, so you pay for maximum reasoning only on the calls that deserve it.
One deflation worth passing on before you budget: the ‘half Fable’s price’ headline doesn’t survive the tokenizer tax. Theo’s measured version is ‘closer to like 20 to 25% off’, because reasoning-token appetites differ. Cheaper, yes. Half, no.
The runners-up
Claude Fable 5 is still the better orchestrator when the brief is genuinely open-ended - see Overall for why - and the best independent evidence says it stays more efficient as runs get very long. If budget allows and the task is vague, it’s gold in practice. It lost the default seat on price, not capability.
GPT-5.6-Sol is enormously capable here, and arguably the best pick of all when the plan is already extremely detailed and was authored by a Fable-class model. Give it a finished PRD and it conducts beautifully. Give it ambiguity and the pit bull interprets your brief the way a genie interprets a wish.
When not to orchestrate at all
New this patch, because someone finally measured it: a company running a 41-task internal benchmark found multi-agent orchestration 2.6× more expensive than just giving the frontier model the work directly - the orchestrator and the worker both pay to read the same context. Orchestration earns its overhead when subtasks are genuinely independent and parallel. If your task is one long dependent chain, the podium above is a list of models to use alone.
What to do now
Default the seat to Opus 5. Escalate to Fable when the judgement calls are still undecided. Hand Sol the baton only when the score is already written - and before you build the orchestra, check your task actually needs more than one musician.