Frontier AI has matured since ChatGPT first arrived a few years ago, but the field still lacks clear boundaries and reliable benchmarks. Developers at the leading edge are chasing two goals at once: greater intelligence and cheaper per-token pricing. The result is a rapidly shifting leaderboard, where some models hold the top spot for only a few hours before being overtaken.
Anthropic’s Claude Fable and Opus models have stayed consistently near the top of the intelligence charts, but they also rank among the most expensive to run. For general use, more attention is turning to the so-called Pareto Frontier, the point where peak intelligence meets minimal cost. Competition there is intense and constantly changing.
After many companies reversed their aggressive “tokenmaxing” strategies earlier this year, the renewed emphasis on lowering AI costs has revived Jevons paradox. As token prices for high-intelligence models have fallen, token usage has climbed more than 25 times over the past year, and has doubled in the past month alone. Users may not want to spend more on AI overall, but they use far more of it when they can afford to.
The Same Names Still Lead the Pack
Despite the industry’s rapid rise, the major players have not changed much. There have been occasional “Deepseek moments” that sent everyone scrambling to respond to a new threat, but the general order holds. OpenAI and Anthropic continue to fight for the top of the intelligence rankings, Google and Meta occupy the efficient, cost-effective middle, and xAI’s Grok remains in the background.
That remains the picture in September 2026. Benchmarks can be gamed during model design, and real-world use tells a fuller story, but Anthropic’s models are still widely regarded as the smartest. Fable 5.1, Fable 5, and Claude Opus all rank in the top four of ArtificialAnalysis’ Intelligence Index, and other testing platforms place them similarly high.
The lead, however, is narrow. Fable 5.1 scores a 66 on the ArtificialAnalysis benchmark, while OpenAI’s GPT 5.6 Sol (max) reaches 61. Grok 4.6 (high) and Kimi K3 (max) both clear 60, and the new Meta Muse Spark 1.3 (max) hits 62, though cost comparison pricing for it is not yet available. Similar patterns appear across other companies’ benchmarks.
Frontier Models Remain Expensive Despite Cuts
Where the leading models trade small advantages back and forth, the real separation appears in the mid-range, driven less by intelligence than by price. Major AI developers acknowledge they have a pricing problem. After the shift to per-token pricing earlier this year, budgets were exhausted, and even AI executives began speaking publicly about making the technology more affordable.
The competitive dynamic increasingly favors mid-tier models, which can deliver roughly 90% of flagship capability at about one-sixth the cost. How aggressive price cutting fits with the industry’s path to profitability is still unresolved, but even after reductions, top-tier frontier models remain costly to operate.
Source
Image: tomshardware.com
