The Brief

New briefs daily around 7 AM Eastern

Test Muse Wonder on the API, but its top scores belong to gated Full

Meta released Muse Wonder on its Model API and a stronger Muse Wonder Full limited to Muse Code subscribers. Independent scores put Full at 56 and Wonder at 52, so API users should benchmark the model they can actually call.

• 01Read

  Artificial Analysis hosts FrontierCode 1.1, scoring PRs on mergeability

  Cognition's FrontierCode 1.1 benchmark now has public results on Artificial Analysis. It grades coding models on whether maintainers would merge their changes, not only on whether tests pass.

  ArtificialAnlys→Cognition→
• 02Read

  TestSuiteBench finds agent test suites kill under 60% of mutants

  Researchers gave coding agents specifications for 165 open-source apps and scored the test suites they wrote by mutation kill rate. Even frontier agents stayed below 60%, although 88% of their tests passed on the reference code.

  Kazuki Egashira, Niels M\"undler-Sasahara, Mark Niklas M\"uller, Martin Vechev→
• 03Read

  Microsoft moves inference to Maia 300 chips at multi-gigawatt scale

  Satya Nadella said Microsoft will scale its Maia chips to multi-gigawatt deployments as its primary inference system, introducing Maia 300 with a claimed 40% more tokens per watt than leading GPU systems.

  Satya Nadella→
• 04Read

  Qwen's QwenGyre cuts long-horizon agent RL training time about 44%

  Qwen introduced QwenGyre, an elastic reinforcement learning system for very long agent tasks. It shifts GPUs between rollout and training and merges branching runs into shared-prefix trajectory trees.

  Alibaba_Qwen→
• 05Read

  ElevenLabs releases Eleven v4 and a Turbo variant near 100ms latency

  ElevenLabs released Eleven v4 text-to-speech with inline delivery tags and a low-latency Turbo variant at about 100ms median latency, now available inside its ElevenAgents platform.

  Eleven V4→Eleven V4 Turbo In Elevenagents→
• 06Read

  HWREBench finds AI agents can hack most cheap smart devices

  A new benchmark, HWREBench, set AI agents against off-the-shelf smart devices and found most could be compromised, starting from an agent that broke a DVD drive's firmware encryption.

  Yenrabbit→