A higher price does not guarantee a smarter model. Among the 31 models with public LMArena scores, we ranked real benchmark Elo against live price.
minimax-m3 scores 1448 on LMArena's human-preference leaderboard while costing only about $0.97 per 1M tokens on a 3:1 blend — within a few points of the very best models at a small fraction of their price. Benchmark Elo is real third-party data, not our estimate.
The highest-scoring model we track, claude-opus-4.6, posts an Elo of 1504 at about $20.00 per 1M. The value leaders land remarkably close on quality while pricing far lower — so for most workloads you are paying a large premium for a small, often unnoticeable, quality difference.
Intelligence per dollar favours models that are both strong and cheap; treat it as a starting filter, not the last word — latency, context window, tool-use and your own evals still matter. Only 31 models currently carry a public LMArena score, so this ranks the measured field, not the whole catalogue. The map below plots price against efficiency.
Every tracked model plotted by input price (log scale) and composite efficiency. Toward the top-left means better value per dollar.
Each dot is one model · color = region · click a dot to open it.
Pricing is real (via OpenRouter, updated daily) and Elo is real (via LMArena). This is market analysis, not investment or procurement advice.