Luiz Sol’s latest analysis applies Modern Portfolio Theory to the chaotic landscape of Large Language Models, identifying GPT-6 Luna as the most efficient model overall. By plotting intelligence scores against cost per task, the study constructs a Pareto frontier that renders 18 of the 26 examined models dominated and effectively irrelevant for cost-conscious developers. The data, retrieved on 1 November 2026 from Artificial Analysis and OpenRouter, suggests that most widely marketed models fail to offer a defensible value proposition compared to the frontier survivors.

The Pareto Frontier of Intelligence and Cost

The analysis defines a model as Pareto efficient only if no other model is both smarter and cheaper. On this strict criterion, only eight models survive: K2 Horizon, GPT-6 Luna, MiMo-V2.6-Pro, Claude Haiku 5.5, GPT-6.1 Terra, GPT-6.1 Sol, GPT-6 Astra, and Gemini 4 Carbon. Notable casualties include Claude Opus 5.5, which scores lower than Gemini 4 Carbon while costing 30% more, and Claude Sonnet 5.5, which offers slightly less intelligence than GPT-6.1 Sol at eight times the cost. This filtering process strips away the marketing noise, leaving a small set of models that actually make economic sense for general-purpose tasks.

Borrowing the Sharpe Ratio for LLMs

To rank the survivors, the author adapts the Sharpe ratio, calculating 'points per dollar' by measuring intelligence gains over the best free baseline. K2 Horizon, an open-weights model from MBZUAI, serves as this risk-free asset with a score of 17.8 and zero cost. GPT-6 Luna emerges as the 'tangency model,' delivering 66 points of excess intelligence per dollar. This metric highlights that Luna’s efficiency stems from its low cost rather than superior raw intelligence, as it clears the free baseline by only 10 points. In contrast, premium models like Gemini 4 Carbon yield a mere 5 points per dollar, indicating diminishing returns at the high end of the market.

Limitations of Financial Analogies in AI

The study explicitly notes where financial theory breaks down in the context of AI inference. Unlike a stock portfolio, a single task cannot be 60% GPT-6 Luna and 40% K2 Horizon; model selection is binary for any given query. Furthermore, the 'risk' in finance implies uncertainty, whereas the cost data here is deterministic. The author acknowledges that real-world costs fluctuate, with coding-agent sessions potentially costing 6 to 11 times more than typical queries due to latency and token usage. This variability suggests that robust optimization, rather than simple average cost, is required for budget-sensitive applications.

Key Takeaways

  • GPT-6 Luna is the most efficient model on cost and intelligence, offering 66 points per dollar over the free baseline.
  • 18 of 26 analyzed models are Pareto dominated, including major players like Claude Opus 5.5 and Grok 4.7.
  • MiMo-V2.6-Pro and Claude Haiku 5.5 are statistically tied for second place in efficiency, differing by only two points per dollar.
  • Premium models like GPT-6 Astra and Gemini 4 Carbon offer minimal incremental value, yielding under 6 points per dollar.
  • The analysis excludes Chinese labs in some scenarios, where GPT-6.1 Terra becomes the top choice for high-intelligence tasks.

The Bottom Line

Stop paying for dominated models. GPT-6 Luna is the clear winner for cost-conscious developers, but remember that real-world latency and session length can skew these theoretical efficiency gains significantly.