Best Value AI Models by LMC ValueScore

A naive score/price ratio explodes as price approaches zero, which is why free models would always dominate a simple division. LMC ValueScore uses a tilted Cobb-Douglas utility index (Q^0.6 times P^0.4) with a smooth sigmoid gate at composite=50. This prevents unproven free models from dominating, and the 60/40 tilt keeps quality slightly more important than price efficiency, which matches what production buyers actually care about.

Early versions of the formula sent premium flagships to LMC ValueScore 0 whenever their price passed the empirical p90 anchor. That produced a PR-disastrous "Claude = 0, Gemini = 82" headline without distinguishing a $4.80/M model from a $140/M one. The final v3 formula applies a soft floor of 1 on the price efficiency component so quality still shows through. Claude Opus 4.6 at composite 90 lands around 15 rather than 0, which is honest: it is clearly not a best-value pick at 8 dollars per million tokens, but it is still one of the highest-quality models available.

LMC ValueScore uses empirical p10 and p90 anchors computed from the priced text-model catalog: p10 = $0.087/M and p90 = $4.80/M as of Q2 2026. We deliberately freeze these quarterly rather than recomputing on every refresh. Morningstar and MSCI rating systems use the same approach: unstable anchors produce unstable rankings that churn for reasons unrelated to the underlying models, which is exactly what SEO and user trust reward us for avoiding.

Yes. Reasoning models bill hidden "thinking" tokens on top of the visible output, so the list output price understates real cost. LMC ValueScore multiplies the output-token contribution by a family-specific expansion factor: o1 at 8x, o1-pro at 15x, o3 at 10x, R1 at 8x, QwQ at 6x, and so on. These factors come from the o1 system card, the DeepSeek R1 paper, and published cookbook traces. That is why R1 lands around 29-37 on LMC ValueScore even though its nominal per-token price looks competitive.

Real chat and RAG workloads are input-heavy: you stuff a lot of context (system prompt, user history, retrieved documents) and get relatively short answers back. Anthropic published caching statistics showing median input-to-output ratios around 5:1 to 10:1 for production traffic, and OpenAI DevDay data agrees. A flat average would bias against reasoning models and against any provider with expensive output tokens. 85/15 is the tightest fit to the median production workload we can defend with public data.