OpenAI's publicly documented work on Astra indicates the AI system has reached a critical cybersecurity capability threshold, scoring 100% on ExploitBench in internal evaluations. The company plans li…
DeepSeek released V4-Flash-Vision-Exp, an experimental multimodal variant of its V4-Flash model, claiming it matches Anthropic's Claude Opus 4.8 on some multimodal benchmarks. The model scores 36.5 on…
A new analysis argues that current AI computer-use agents fail to generalize to real-world interfaces because they bypass the UI entirely, relying on API calls and scripting instead of human-level int…
UC Berkeley RDI's Agents' Last Exam (ALE) benchmark, released in June 2026, found that every tested frontier AI agent scored 0% on its hardest tier of long-horizon professional tasks, while PYMNTS rep…
OpenAI shipped GPT-5.6 on July 9 as three distinct models — Sol, Terra, and Luna — not a single upgrade, with the gpt-5.6 alias routing to Sol at $5 per million input tokens. The release introduces Pr…
OpenAI released the GPT-5.6 family of models in three sizes—Luna, Terra, and Sol—claiming superior long-running agentic performance over Anthropic's Claude Fable 5 on the Agents' Last Exam benchmark, …
Meta Platforms' new vice-president of AI research, Dawn Song, said the next frontier of artificial intelligence will be agents that can perform economically valuable work across real-world domains. So…
A new wave of long-horizon agent benchmarks, including Agents' Last Exam, SWE-Marathon, the Meta-Agent Challenge, and Arena's Agent Mode, reveals a fundamental trade-off between realism and verifiabil…
Top AI agents scored 0% on expert-level professional tasks in the ALE benchmark, with models like Fable 5 and GPT-5.5 failing to solve any problems requiring real expertise. Mid-level performance reac…
Berkeley RDI and over 300 industry experts launched Agents' Last Exam (ALE), an open benchmark that evaluates AI agents on economically valuable professional tasks rather than abstract proxies. The be…
Researchers at MiniMax introduced MiniMax Sparse Attention (MSA), a method that reduces attention compute by 28.4x at 1M context while matching performance, enabling cheaper long-context deployment. B…
Anthropic launched Fable 5 / Mythos 5, a model tier for long-horizon asynchronous autonomy, on June 10, 2026, marking the arrival of the async-agent era. The release reframes AI competition around aut…
Researchers introduced Agents' Last Exam (ALE), a new benchmark designed to evaluate AI agents on long-horizon, economically valuable, real-world tasks with verifiable outcomes. Developed with over 25…