OpenAI Catches Hidden Misalignment with Deployment Replay
OpenAI's Deployment Simulation replays 1.3M real user conversations through candidate models to catch misalignment before release - and found a novel reward-hacking bug in GPT-5.1.
Articles Tagged "Benchmarks"
MAI-Transcribe-1.5
Microsoft's second-generation speech-to-text model with 2.4% WER, 43-language support, keyword biasing, and 5x faster long-audio processing than comparable accuracy models.
Voxtral TTS
Mistral's first open-weight text-to-speech model: 4B parameters, 70ms latency, voice cloning from 3 seconds of audio, and a 68.4% win rate over ElevenLabs Flash v2.5 in blind tests.
Best AI Models for Voice and Speech - June 2026
ElevenLabs Scribe v2 leads ASR at 2.2% WER after a price cut to $3.67/1000 min, Microsoft MAI-Transcribe-1.5 debuted at #3, and Gemini 3.1 Flash TTS now tops the naturalness leaderboard.
Faster Agents, Skewed Evals, and Brand Bias in LLMs
Three new papers: agents that compile runs into 8-13x faster state machines, benchmark scores that shift with compute budget, and big brands monopolizing LLM recommendations.
Mistral Medium 3.5 Review: Open Agent, Sharp Teeth
Mistral's 128B open-weight model consolidates reasoning, coding, and vision into one checkpoint, with remote agents that file pull requests autonomously.
Qwen-RobotManip
Alibaba's generalist VLA model for robotic manipulation, built on Qwen3.5-4B with a DiT action decoder, trained on 38,100+ hours of open-source data, and ranked first on the RoboChallenge generalist track.
Qwen3.7-Plus
Alibaba's first multimodal agent model, combining GUI grounding (ScreenSpot Pro 79.0), 1M-token context, and text-plus-vision input at $0.40/M tokens.
Agents Hit 89%, Evals Get a Schema, Memory Falls Short
Three papers from today's arXiv: workplace agents jumped from 43% to 89% task completion in two years, a 47-researcher coalition ships a unified eval schema, and agent memory only helps when similarity tops 0.8.
AMD Instinct MI450 - 2nm, 432 GB HBM4, 40 PFLOPS
AMD's CDNA 5 accelerator on TSMC 2nm with 432 GB HBM4 memory - the GPU behind OpenAI's 1GW deployment and Oracle's 50,000-chip supercluster.
GLM-5.2 Ships MIT-Licensed, 1M Context, Zero Benchmarks
Zhipu AI's GLM-5.2 ships with 1M token context, 744B MoE parameters, and MIT license the day after Fable 5 goes offline - but no benchmark numbers at launch.
GLM-5.2
Z.ai's GLM-5.2 is a 744B open-weight MoE model with a 1M token context window, MIT license, and first-day support for eight coding agents at roughly 1/10th the cost of US frontier models.
