Efficiencymaxxing: optimizing the AI intelligence-to-cost curve

There’s a moment going on right now around AI cost efficiency, and I’m here for it. For the past year, most teams using AI have optimized for maximum intelligence per request. The next phase is optimizing for the right mix of intelligence, speed, and cost. Efficiencymaxxing instead of Tokenmaxxing.

The curve is expanding

Lots to be excited about from last week that aim to expand the area under the intelligence-to-cost curve:

• Moonshot launched Kimi K3 ($3 / $15), an open-weight model that benchmarks competitively against the frontier labs’ high-end models.
• OpenAI slashed pricing on GPT-5.6: Luna by 80% and Terra by 20%.
• DeepSeek launched V4 Flash 0731, an open-weight model that competes with the frontier labs’ mid-tier models at wildly efficient pricing ($0.14 / $0.28).

My efficiencymaxxing experiment

I’ve spent the last few days trying to make my own AI usage in Adapt more efficient. I wanted to be a guinea pig before we considered providing customers with non-frontier models within Adapt. The early results are incredibly encouraging.

Last week I was running a model recipe of Anthropic’s Sonnet 5 ($2 / $10), Opus 5 ($5 / $25), and Fable 5 ($10 / $50). Adaptive Intelligence dynamically selected the right model from that stack based on what question I was asking. This week, I switched that recipe over to a mix of OpenAI’s GPT-5.6 Luna ($0.20 / $1.20) and GPT-5.6 Terra ($2 / $12). A quick lookup or one-liner went to Luna. A long report with a dozen tool calls went to Terra at high effort.

Comparing Monday and Tuesday on my efficient OpenAI recipe against the last three Mondays and Tuesdays on my Anthropic recipe, my average cost per turn is down 83%. Adapt set up its own rubric to rate its responses, and any degradation in quality from switching to a more efficient recipe is negligible. My spend at list API rates was closer to $10 per day than $100 per day, a major difference. I was able to continue working at the same pace and level while banking the kind of spend decrease that fires up our finance team (me).

Match the model to the task

I’ve been really excited to be the guinea pig on this. The point is to match the task at hand with the right level of intelligence and effort. What someone considers to be the “best” model is subjective, but it sits somewhere on an intelligence-to-cost curve. The best model is not always the most intelligent model available, and it is not always the cheapest one either.

With Adapt, our users can choose the models that match what they think is best for the work. No re-platforming, no model migration project, no one-lab lock-in.

Our goal is to help you spend up to the point of diminishing returns on AI, and not a penny more. This is us eating our own cooking.

I won’t be returning to the frontier-level recipe after this experiment. My ROI is far better at $10 per day than nearly $100, as long as my output quality stays high.

What I’m testing next

Over the next few days I’ll test a mix of DeepSeek’s V4 Flash 0731 ($0.14 / $0.28) and Moonshot’s Kimi K3 ($3 / $15) as the frontier-competitive end of the recipe. The fact that both are open-weight gives us more flexibility around how we could serve them to customers. I’m excited to see what this recipe can do and whether I can push spend down further. I’ll report back.