Frontier LLMs · August 2026
Every frontier model, plotted by how much a unit of intelligence actually costs against how capable it is. The horizontal axis is the real bill — the cost of one Artificial Analysis Intelligence Index task, reasoning tokens included. Open weights still own the cheap end of the curve, but Kimi K3 has lost the rung it opened in July: OpenAI's 21 Aug price cut on GPT-5.6 Sol made it both cheaper and higher-scoring than Kimi K3 at once. Xiaomi's MiMo-V2.5-Pro remains the best value on the entire chart, at nearly 1,000 index points per task-dollar. Hover any point for its exact coordinates; toggle the axis to see what a dollar of task-cost really buys.
Vertical: Artificial Analysis Intelligence Index. Horizontal: cost (USD) per Intelligence Index task, log scale — lower is better. Scores and costs read from Artificial Analysis model pages on 21 Aug 2026; GPT-5.6 Sol's figures are current to 25 Aug 2026 and cross-checked against OpenAI's own pricing page.
Method: x = each model's published cost to run the full Intelligence Index, divided by ~2,200 tasks. The divisor is an estimate, so absolute x-values are approximate — but it is the same divisor for every model, so the comparisons between them are not. AA has rebased the Index, so these scores are not comparable with figures quoted in its older launch articles.
Not plotted: Gemini 3 Pro (index 41), for which AA reports no evaluation cost, and Nemotron 3 Ultra, which no longer appears in Artificial Analysis's tracked models. Qwen3.8-Max is marked paid — Alibaba has promised weights but not yet shipped them.
Artificial Analysis model index First-hand benchmark results →
The dashed line is the efficient frontier — the most intelligence available at each cost-per-task. Eight models hold it: gpt-oss-120b and MiMo-V2.5-Pro at the cheap end, then GPT-5.6 Luna, Grok 4.5, GPT-5.6 Terra, Qwen3.8-Max, GPT-5.6 Sol and Claude Opus 5 at the ceiling. Two of those eight are open weights — and they hold the two cheapest rungs outright, not just a foothold.
OpenAI cut GPT-5.6 Sol's price on 21 Aug 2026, from $5.00/$30.00 to $4.00/$20.00 per million input/output tokens, and Artificial Analysis's evaluation cost for Sol fell with it, from about $1.28 to about $0.92 a task, without moving its index (61). The cut makes Sol strictly better than Kimi K3 for the first time — cheaper AND higher-scoring — which pushes Kimi K3 off the frontier it opened in July. The best-value point on the whole chart, paid or open, remains Xiaomi's MiMo-V2.5-Pro: index 43 for about $0.045 a task, roughly 957 index points per task-dollar — nearly 3x DeepSeek V4 Flash's number, and open weights again.
Kimi K3 is fourth in the world by raw Intelligence Index (60), but it holds no rung of its own on the frontier — at about $1.10 a task, GPT-5.6 Sol undercuts it while scoring higher. GPT-5.6 Luna is the best value of anything closed — index 52 for about $0.08 a task, roughly 664 index points per task-dollar — but it's not the best value on the chart; MiMo-V2.5-Pro beats it outright. Claude Fable 5 sits at the other extreme: index 62 for about $2.48 a task, roughly 25 points per task-dollar.