# Route AI Models by Temperament, Not Just Cost Tier Two frontier models at comparable capability and price can still have opposite *temperaments* — stable behavioral dispositions that decide how they fail, not whether they're smart. So the routing question isn't only "which tier can afford this task?" but "which model's failure mode can this task tolerate?" One model reads instructions literally and, when it hits the edge of its knowledge, invents a plausible answer to keep moving. Another reads for intent and, at that same edge, stops and admits it doesn't know. Neither is better in the abstract; each is right for different work. <!-- atomicity: 1 idea ✓ --> ## Why This Matters Theo Browne's Fable-5-vs-GPT-5.6 comparison makes the axis concrete. At rough capability parity, GPT-5.6 "reads what you said and does exactly that" and is "very quick to work around it or make something up" when it doesn't know; Fable "reads between the lines" to infer intent and "is very quick to admit" ignorance. That yields opposite operating instructions: the literal model needs a *complete, unambiguous* spec (it won't rescue an underspecified prompt, but it also won't wander), while the intent-inferring model needs *guardrails against over-reach* (it finds better solutions but can drop stated details in pursuit of the goal). The tolerance question follows directly: for a task where a confidently-wrong answer is expensive to catch (security findings, migrations, anything you'll merge without deep review), route to the model that admits ignorance; for a task where you've written an exact spec and want it executed without editorializing, route to the literal one. This is a *different axis* from cost-tier routing. Cost routing asks "is this worth the expensive model?" and sends judgment to the strong tier and bulk to the cheap tier. Temperament routing operates *within* a tier: given two models you can equally afford, pick by disposition. The two compose — first bound by cost, then choose by temperament — and a mature routing policy uses both. ## Scope & Conditions - **Holds** when you have two or more models of similar capability/price available and real latitude to choose per task (agent frameworks, multi-model subscriptions, API routing layers). - **Requires** that you actually know each model's temperament — which means empirical use, not benchmark scores. Benchmarks measure capability (can it?), not disposition (how does it fail?); two models can tie on a benchmark and diverge completely here. Behavioral claims like these are also model-version-specific and decay as models update — re-derive them, don't inherit them. - **Breaks down** when only one model is viable (capability floor, cost ceiling, or availability), or when the task is trivial enough that failure mode doesn't matter. ## Links - [[3. Resources/Permanent Notes/Outcome-Based Pricing Transfers AI Cost Risk From Buyer to Vendor]] — applies-in: outcome pricing gives the vendor the incentive to route each task to the cheapest model temperament that still meets the promised result. - [[3. Resources/Permanent Notes/Route Judgment to the Expensive Model and Bulk Execution to the Cheap One]] — the sibling axis: that note routes by cost/capability tier, this one routes by temperament *within* a tier; a full policy composes both - [[3. Resources/Permanent Notes/The Bottleneck Moved From Code Generation to Review and Judgment]] — why temperament matters now: when review is the binding constraint, a model that admits ignorance lowers review cost more than a marginally-smarter one that hides its uncertainty - [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]] — the failure-mode-tolerance framing depends on independent verification; a model that "makes something up to unblock itself" is exactly why the checker can't be the maker - [[3. Resources/Literature Notes/Fable 5 vs GPT-5.6]] — the source comparison and the owl-vs-Rottweiler mental model ## Source - [[3. Resources/Literature Notes/Fable 5 vs GPT-5.6]] — Theo Browne (t3.gg), 2026: the temperament contrast (literal/invents vs infers/admits), the "brutalize it into listening" vs "brutalize it into not over-reaching" prompting split, and the Omniscience-benchmark grounding for admits-vs-invents. Reader doc 01kxtjnrt3hd0xjyyex7gdez0s. (Benchmark figures self-reported; verify against Artificial Analysis.) - See also: [[3. Resources/MOCs/AI Cost & Tokenomics MOC]]