Agnes 2.5 Pro Alpha

Agnes AI, a Singapore-based lab that trains models in-house, announced Agnes 2.5 Pro Alpha as the first Agnes AI text model to appear on the Artificial Analysis leaderboard, where its image and video models already rank in the top ten. The announcement, distributed via Business Wire and carried by Fundatio Spelling, fr According to the release, on Artificial Analysis Intelligence Index v4.1 (a composite of nine evaluations) Agnes 2.5 Pro Alpha scored 39, ahead of Cohere's Command A+ (23) and Mistral's Devstral 2 (19) in its cohort. It reports 67% on Terminal-Bench v2.1, ahead of DeepSeek V4 Pro, and 88% on GPQA Diamond, near the list

Agnes 2.5 Pro Alpha, the first Agnes AI text model to appear on the Artificial Analysis leaderboard, debuted with an Intelligence Index v4.1 score of 39, placing it ahead of Cohere's Command A+ (23) and Mistral's Devstral 2 (19). The Singapore-based company trained the model in-house, and its image and video models alr On technical benchmarks, Agnes 2.5 Pro Alpha scored 67% on Terminal-Bench v2.1 (running commands end-to-end, ahead of DeepSeek V4 Pro) and 88% on GPQA Diamond, close to the leading 94%. The press release highlights efficiency: average tasks on GDPval-AA v2 cost roughly $0.035 (10,000 monthly tasks ≈ $342, versus $12,30

Agnes AI has released Agnes 2.5 Pro Alpha, a reasoning model trained in-house by the Singapore-based lab and offered through its omni-modal API. According to Artificial Analysis, the model accepts text, image, and video inputs and produces text output, with a 1M-token context window and pricing of $0.45 per 1M input to The model's standout trait is price-to-intelligence positioning: at the 38–40 Intelligence Index band, only DeepSeek V4 Flash (Reasoning, Max Effort) is cheaper on a 7:2:1 blended basis ($0.06 vs $0.18 per 1M tokens), and open-weights alternatives like DeepSeek V4 Pro and MiMo-V2.5-Pro match the $0.18 blended price at

BenchLeader's profile for Agnes 2.5 Pro Alpha confirms the model's benchmarkable status and provides an independent, third-party view of its evaluation results and runtime costs. Reported scores include GPQA Diamond at 87.6%, Humanity's Last Exam at 33.6%, SciCode at 42.9%, AA-Omniscience at -25.1, AA-LCR at 75.7%, and The page also models latency and cost across five realistic workloads using list pricing and output speed: a 400/300-token chat reply costs about $0.0005 in roughly 17.4 seconds; summarising a 30-page report (12,000/600 tokens) costs about $0.0059 in 19.1 seconds; a code edit (6,000/1,500) runs about $0.0041 in 24.4 se