When to Escalate From the Small Model: Watch How It Fails
When to Escalate From the Small Model: Watch How It Fails
In the Fronset benchmark’s model portfolio, every activity gets a default model and, only where the evidence supports one, a fallback. The small model can sit in either slot, and on several activities the fallback is left empty. What should move a call from the default to the fallback depends on how the default fails, not only on how well it scores.
Which model takes the call first
Where the small model’s score is tied with the large one’s, the portfolio makes the small model the default and the large one the fallback, as with Claude Haiku 5.5 and Claude Opus 5.5 on capability category matching [^1] and GLM-5.3 Flash and GLM-5.3 on checking whether a persona meets a policy [^2].
The order reverses where the large model leads. On selecting authoritative sources, GLM-5.3 scores 8.65 (±0.15) against 8.11 (±0.29) for GLM-5.3 Flash [^3], where ± is the confidence interval, the margin of uncertainty around each score. GLM-5.3 takes the call first and the Flash model is the fallback.
The fallback does not have to be a different model. On configuring a visual theme, Claude Opus 5.5 scores 8.72 at $0.04 in batch mode and $0.09 in sync mode [^4]; the score is identical and only the price moves. The portfolio makes batch the default and sync the fallback for the same model.
When there is nowhere to escalate
On several activities the portfolio names a default and no fallback. For two of them the reason is stated: nothing else measured clears its quality bar of 8.0 out of 10. Claude Sonnet 5.5 writes promotional messages, and Claude Haiku 5.5 reaches 7.76 (±0.26) on the same work [^5]. NVIDIA Nemotron-3 Ultra 550B rewrites metadata paragraphs, and the tied low-cost cluster below it tops out at 7.9 (±0.17) for DeepSeek V4.1 Flash [^6].
On these routes a failed call has no model to go to next. On claim continuity classification the portfolio goes a step further and places its default, Claude Opus 5.5, behind a review step [^7], so the backstop there is a check rather than another model.
What should trigger the fallback
Escalation needs a signal, and the type of failure decides whether there is one. On claim continuity classification, Qwen 3.7 Plus fails mostly by terse output, which is visible, while Meta Muse Spark 1.3 fails mostly by assigning the wrong class, which may pass unnoticed; their judged failure rates on that work, meaning the share of outputs the benchmark graded as failures, are 39.5% and 27.5% respectively [^7]. A visible failure can be caught and sent on to the fallback. A wrong-class failure looks like a normal answer, so where such failures would go unseen, the route needs a check.
Output that cannot be parsed is easy to detect, and in this material it is not cheap. On promotional campaign generation, $39.70 of $61.27 spent bought output that could not be parsed [^8], though those losses are not attributed to models. Thinking Machines Inkling Small needed a separately billed repair call on 2.9% of 9,639 attempts [^9], so a repair or fallback path carries a cost of its own.
Setting up the routes
Put together, the portfolio suggests a simple structure for each activity:
- Assign a default and a fallback per activity, and leave the fallback empty where nothing measured clears your bar, rather than routing to a model the measurements do not cover.
- Escalate on failures you can detect, such as terse or unparseable output, and put a review step behind any default whose failures pass unnoticed.
- Count repair and fallback calls in the budget, since they are billed on top of the default call.
- Revisit routes decided by an overlapping interval at each snapshot, because those are the ones most likely to flip.
The live benchmark at https://fronset.ai/benchmark/ carries these measurements and the models behind them. Set the quality bar your production output requires there before you fix a route.