Tag: benchmarks
Four Gemini agents finish Minecraft with no humans involved
A DeepMind team dropped four Gemini 4 Argon agents into a random seed and they killed the Ender…
Grok Voice 3.0 tops a speech benchmark ahead of OpenAI
SpaceXSI's new voice model leads on naturalness while costing less per call task than its closest rivals.
Stripped guardrails let open models answer terror prompts
A new benchmark found that removed safety layers turn widely available open-weight models into willing assistants for extremist…
Sakana lets one model grade and improve its own work
A single language model that solves, judges and refines its own workflows posted real gains, but shared blind…
Agent teams burn far more tokens for almost no gain
A benchmark of GPT-6 Sol and Claude Opus 5.5 finds teams beat solo agents in just one of…
Arena doubles its valuation to $3.1B on model evaluation demand
The Berkeley-born AI leaderboard Arena has raised a $200M Series B at a $3.1B valuation, nearly double its…
Nvidia’s Kumo Tabular predicts table labels in one pass
An open NVIDIA foundation model for tabular data skips training, tuning and feature engineering.
H Company trains one agent model to click, code and call APIs
Holo4 ships in two sizes and switches between screen control, code execution and tool calls without swapping models.
Exa opens an ultra deep research mode that swarms subagents
The search startup's highest effort setting spends up to three hours on a single research request.
An open model sorts eight overlapping speakers in real time
NVIDIA released Nemotron 3 Diarization, a small open model that tops VoiceArena's leaderboard while running live conversation streams.
CheatBench scores how often agents cut corners on hard tasks
The Center for AI Safety tested frontier agents on tasks with hidden ways to cheat, and every one…
Robot arms rarely refuse in a new harm benchmark
Given instructions to cause harm with real hardware, three leading models almost never said no.