Nvidia has demonstrated a perfect score of 100% on the ARC AGI-3 benchmark using its novel harness, AVO. The article notes that OpenAI did not use this standard harness for their results on the same benchmark.
| Benchmark | Model | Score |
|---|---|---|
| ARC-AGI 3 | Nvidia AVO | 100% |
NVIDIA has released its AVO architecture, which achieves a perfect score on the ARC-AGI-3 benchmark. The system is designed to support long-horizon autonomous agents by managing context, tool usage, and state recovery.
OpenAI has launched GPT-5.6, introducing a new model ladder of Luna, Terra, and Sol alongside distinct effort levels for Max and Ultra modes. The release brings significant changes to the ChatGPT Work and Codex interface, though it initially faced user backlash regarding confusing navigation and rapid usage depletion.
A study evaluates four AI review systems across six language models, finding OpenAIReview with GPT-5.5 achieves 83.0% accuracy in matching paper quality to external signals and detects 71.6% of injected errors. Real user feedback shows positive sentiment, with a 1.44-to-1 vote ratio, though false positives and minor nitpicks remain common.
Artificial Analysis' new agentic knowledge work evaluation, AA-Briefcase, shows GLM-5.2 surpassing GPT-5.5 in performance. The benchmark assesses real-world task execution and reasoning capabilities in knowledge work scenarios.
TokenPilot reduces inference costs by 61% to 87% in both isolated and continuous modes, outperforming prior systems in cost efficiency while maintaining competitive performance. It uses ingestion-aware compaction and lifecycle-aware eviction to stabilize prompt prefixes and manage context segments efficiently.
We use cookies to measure traffic and improve the site. You can accept or decline analytics cookies. Privacy policy