Incentives work, but not because they are money
Paying a mid-tier model for correct answers lifts accuracy from 57% to 76%. Telling it a life depends on the answer works slightly better. Stakes are a framing channel, not an economic one.
- 57.3% → 76.0%
- Sonnet accuracy, baseline vs. incentivized (1,200 calls per model)
- +46pp
- effect at the capability boundary (hard tasks, p < 0.001)
- 79%
- accuracy under the non-economic control: a life depends on this
- $4.19 vs. $8.66
- cost per 1,000 tasks, incentivized Sonnet vs. Opus baseline
The setup
A single agent solves procedurally generated math tasks across four difficulty levels, 1,200 calls per model, every answer verified deterministically with numpy. No LLM judges. Conditions: baseline, a token bonus for correct answers, a penalty for wrong ones, and both combined.
The effect lives at the capability boundary
Incentivized Sonnet jumps from 57.3% to 76.0%, close to the Opus baseline of 78.7% at about half the cost. The effect concentrates where the model struggles but is not overwhelmed: nothing on easy tasks (ceiling), +19 points on medium, +46 points on hard (p < 0.001), and no significant effect on very hard tasks (floor). Incentives do not create capability. They unlock it at the boundary.
The twist
The control conditions break the economic story. “A human life depends on this” reaches 79%. Forced chain-of-thought reaches 73%. A padded prompt of the same length with no stakes stays at 55%. Money is one high-stakes framing among several, not a special channel.
Implications
For builders: before reaching for a bigger model on mid-hard tasks, try raising the stakes in the prompt. It is a much cheaper lever.
For research: testing genuinely economic mechanisms requires real state, budgets, costs, consequences. Prompt decoration only ever tests framing. That conclusion set the design constraint for everything we built afterwards.
Limits: one model family, math tasks, prompt-level stakes without persistent state.