Jiantao Jiao says an NVIDIA team built SWE-Serve to test whether AI agents can develop an inference engine and serve real models. The benchmark has 53 tasks drawn from real sgl_project engineering work.
WebCraftBench has agents use a live app, then scores aesthetics, usability, and whether the original request was met. Tencent Hy reports an 85.3% match with human preference on 197 validated pairs.
OpenAI says MentalHealthBench is designed for the full range of mental health conversations people bring to AI, from everyday support to more acute crisis scenarios, rather than only emergencies.
OpenAI has openly released MentalHealthBench, a new benchmark built with input from more than 80 mental health clinicians. OpenAI says it shows frontier models continuing to improve in realistic mental health conversations.
On Artificial Analysis’s Pronunciation Robustness benchmark, Gemini 3.8 Flash TTS scores 89.5%, ahead of Gemini 3.1 Flash TTS, SpaceXAI TTS, and Gemini 3.8 Flash-Lite TTS. Leadership on individual pronunciation categories is split among several models.
Polymarket says GPT-6 Astra was the only AI tested by DrivingBench to complete the course.
NVIDIA AI reports a 14.72% error rate for Nemotron 3 Diarization on voicearena_ai’s initial Diarization-Bench, about 24% lower than the runner-up.
Tarush Agarwal reported a speech-to-text benchmark of 15 models on the same audio and pipeline. On 1,000 read-aloud clips from Pipecat FLEURS, AssemblyAI Universal 3.5 Pro had a 1.93% word error rate and a 489 ms median time to first text, both the best results in that test.
Lovable says GPT-6 Sol scores 6 to 12% higher than GPT-5.6 Sol at every effort level on its 0-to-1 building benchmark, with larger gains in intent alignment and design fundamentals.
ModelScope said Qwen has introduced RecreationBench, a 250-task benchmark for Hybrid Computer-Use Agents across Ubuntu, macOS, Windows, Android, and Web. Agents must explore a running app, recreate it in code, and pass programmatic tests plus VLM-based visual evaluation.
In a quick demo, TypeSafe AI's Jev scored 80.5% exact at about 200 milliseconds and $122 per 1M, against Luna's 84.9% exact at about 2 seconds and $482 per 1M.
