“You can’t use effort to pop you up to a higher level of reasoning. The more performant models are going to be more performant. The effort just means how many times is it going to go through the decision tree before it gives up, or how many responses does it get before it says ‘good enough.’” — Lou
Session context: 2026-06-18_Mastermind — Lou re-ran his Model Selection Playbook charts (Fable removed) to answer one question: can you crank a weaker model’s effort high enough to match a stronger model’s reasoning?
Core Idea
The answer is no, and the distinction is the whole insight. Reasoning capability is a fixed property of the model. Effort is a separate dial that controls depth of exploration, not depth of thought. Raising effort makes a model traverse more of the decision/exploration tree before it accepts an answer — it doesn’t grant the model reasoning it doesn’t have. Sonnet at max effort is still Sonnet; it explores harder, but its ceiling is unchanged. So “high effort on a cheaper model” is not a substitute for a more capable model — it’s a way to make a capable-enough model dig deeper.
The cost consequence is the trap. Lou spent a night pushing Sonnet 4.6 through a coding task at high effort and “blew through all my subscription fees plus another $75 in extra usage.” By the time you bump effort and absorb all the extra loops a weaker model takes to converge, “it costs as much, if not more, than just using Opus in the first place.” Effort isn’t free; it’s billed in iterations.
The second finding is that the gap is task-dependent. For coding, Sonnet is “considerably less capable” than Opus — the reasoning ceiling bites hard. But for creative, advisory, and sales-style writing, Sonnet is “much more performant relative to Opus” — close enough that the ceiling rarely matters. That asymmetry is the actual routing rule: reserve the flagship model for work where the reasoning ceiling is load-bearing (code, deep strategy, anything bylined), and let the cheaper model carry the everyday work where it isn’t.
Practical Application
Before you reach for the effort slider on a cheaper model, ask: is the failure I’m seeing a depth problem or a ceiling problem? If the model just needs to explore more (it accepted a shallow first answer), raise effort. If the model literally can’t reason its way to the answer, raising effort only burns iterations — switch models instead. Lou’s standing calibration: highest model you can at medium-to-high effort for what counts (code, strategy, anything with your name on it); Sonnet for everyday work (newsletters, drafts, descriptive content); and never use Haiku Max where Sonnet Low would do — “Sonnet wins by a mile.”
Related Insights
- Insight - Model Altitude — Route Model and Effort by Workflow Step, Not by Whole Artifact — the per-step routing framework; this insight supplies the hard constraint that keeps the effort dial honest.
- Insight - Control AI Reasoning Effort to Stop Context Pollution — establishes effort as a controllable parameter; this sharpens what effort actually buys (depth, not reasoning).
- Insight - Choose Your Claude Model by Task Type, Not by Default — the task-dependent gap (code vs. creative) is the routing rule that follows from the ceiling.
- Insight - The Platform Loyalty Principle — Don’t Platform-Hop When AI Models Are Leapfrogging — the cost discipline behind both.
Evolution Across Sessions
Builds directly on Insight - Model Altitude — Route Model and Effort by Workflow Step, Not by Whole Artifact (2026-06-11), which said route model and effort per step. The new development is empirical and corrective: effort and reasoning are orthogonal, not interchangeable. You cannot substitute effort for capability — and trying to often costs more than the capable model would have. This converts “model altitude” from a preference into a constraint.