Token efficiency

Nvidia

2026-09-26 10:37:29

Nvidia says SoL-Pi cuts coding agent token use by as much as 49%

Nvidia’s SoL-Pi system reduces token consumption for coding agents by optimizing the control layer between the model and its environment, according to a Techub report citing The Decoder. The reported reduction reaches as much as 49%, while performance changes remain limited. The system was developed by a research agent using 152 methods across more than 3,000 tests. The report also noted that gains were smaller on other benchmark tests. The update centers on efficiency rather than a broad jump in benchmark performance, based on the information provided.

20

GPT-6 Astra

2026-09-10 10:05:37

GPT-6 Astra posts first perfect score on Korean CSAT-style AI test, beating GPT-5.6 by 1.5 points

GPT-6 Astra scored a perfect 450 on a Korean College Scholastic Ability Test-style evaluation for large language models, becoming the first model to finish the full exam without losing a point. According to The Korea Daily, Astra earned full marks across Korean language, math, English, Korean history, and selected subjects including Physics I, Chemistry I, Biology I, and Social Studies and Culture. The runner-up, GPT-5.6, scored 448.5, leaving a gap of just 1.5 points. The score gap was narrow, but the compute gap was more pronounced. Astra used 357,000 tokens to complete the exam, versus 429,000 for GPT-5.6 and 562,000 for Claude Fable 5.1. That put Astra about 17% below the second-place model in token usage and 36% below Claude Fable 5.1. The test was run on GitHub using the actual 2026 academic year CSAT questions administered in 2025, with external web search barred during the evaluation. The report also cited professor Lee Seung-hyun, who said Astra’s edge appeared to come from structural efficiency in how it compressed and revised its solution plan while preserving its reasoning flow.

740

Google DeepMi

2026-09-01 17:34:18

Google DeepMind Introduces Agentic Video Understanding for Gemini, Cutting Token Usage by Up to 88%

Google DeepMind has announced the addition of agentic video understanding capabilities to its latest Gemini model, enabling higher accuracy in video analysis while reducing token consumption by up to 88%. The feature, announced via the company's official channels, marks a significant efficiency improvement for video processing tasks, potentially lowering computational costs for developers. The agentic approach optimizes how the model processes video frames, though specific technical details were not disclosed. This advancement continues DeepMind's focus on enhancing multimodal AI efficiency.

230