Two models in our tracked catalog score 42 on the Artificial Analysis Intelligence Index v4.3.
Anthropic's Claude Opus 4.8 shipped 27 May 2026 at $25.00 per million output tokens. Z.ai's GLM-5.3 Flash shipped 26 August 2026 at $0.50.
Ninety-one days. Same index score. Fifty times the output price.
The two, side by side
Read the last row before drawing conclusions. GLM-5.3 Flash is 43% slower on median output throughput. A composite index score is not a claim that two models behave identically, and anyone swapping one for the other on the strength of a single number will find the differences the index does not measure.
The trend behind the pair
Across the 39 models in our catalog that carry both an Intelligence Index score and a published output price, the median cost per index point has roughly halved in two months:
The extremes are wider still. Claude Fable 5 costs $1.00 per index point; InclusionAI's Ling 3.0 Flash VL costs $0.0072, a spread of about 139x across models that are simultaneously on sale.
Where the collapse has not reached
The top of the index. Of the 39 models, nine score 40 or better, three reach 45, and two reach 50. The cheapest model at 45-plus is OpenAI's GPT-5.6 Sol at $10.00 per million output; the cheapest at 50-plus is Claude Opus 5 at $25.00.
So the honest summary is not "frontier intelligence got cheap". It is that the capability band that was frontier in May is now commodity, while the band that is frontier today still prices like it. If your workload needed an index score of 42 in May, your bill can fall by a factor of fifty. If it needs 50, it cannot.
The caveat that matters most
The Intelligence Index is a composite of ten evaluations. It is a useful ordering signal and a poor substitute for testing your own workload. Two models at 42 can differ sharply on the specific thing you do — and on throughput, latency, context handling and refusal behaviour, none of which the index scores.
Index scores are from each model's Artificial Analysis page, which states the index version alongside the number. Prices, context and throughput are from each model author's own served endpoint. Every figure carries its source and date on the model pages.
