When cross-vendor orchestration isn't worth it

I wrote about orchestrating ChatGPT and Gemini from Claude Fable 5 two weeks ago, and the response was mostly some version of so I should always run three models now? No. Here's when it isn't worth it, learned the expensive way.

Latency compounds if you're not careful

Three sequential API calls at two to four seconds each is a ten-second round trip before the orchestrator even starts writing its synthesis. I run the two subordinate calls in parallel, which gets it back to roughly one model's latency plus overhead, but plenty of orchestration frameworks default to serial tool calls, and nobody notices until a reviewer is waiting fifteen seconds for what was supposed to be a quick sanity check.

Failures don't fail independently

The whole pitch for ensemble review is that GPT-5.6, Gemini and Fable 5 make uncorrelated mistakes. Fine for surface bugs, off-by-one errors, a wrong regex, a missed null check, but not for anything downstream of a shared training-data gap: all three models will confidently misuse the same obscure library API if none of them saw good examples of it during training, and three confident wrong answers agreeing with each other is worse than one because it reads as consensus.

• Ensemble review catches implementation slips, not knowledge gaps shared across vendors
• If your team keeps hitting the same wrong answer from all three models, that's a documentation problem, not an orchestration problem
• Check the actual library docs yourself before trusting three-way agreement on anything niche

Cost multiplies, budget for it

A three-model pass costs three to four times a single pass in tokens, and if you're running this on every pull request rather than just the risky ones, that bill adds up faster than the caching tricks in the architecture that cuts 99% of your LLM bill can offset. I gate it behind a label, needs-cross-review, rather than running it by default.

One vendor going down takes the whole pipeline with it

Ensemble review is only as reliable as its least reliable API.

OpenAI, Google and Anthropic each have their own status pages and their own outage schedules, and those schedules aren't correlated either. Build a real timeout and fallback: if GPT-5.6 doesn't answer in eight seconds, proceed with the two you have rather than blocking the whole review on one slow vendor. I learned this after a Gemini rate-limit spike silently stalled a CI job for eleven minutes because the tool call had no timeout set at all.

The actual bar

Reserve cross-vendor orchestration for the reviews where being wrong is expensive: auth changes, payment code, data migrations, anything touching PII. For the rest, one strong model run well, see agentic architectures that don't fall over, beats three models run carelessly, every time.