AI benchmark scores can mislead when used for critical decisions. This guide outlines seven essential checks—from data contamination to reproducibility and model cheating—for determining whether a benchmark result is trustworthy.
MLCommons has announced the release of MLPerf Client v2.0, the latest version of its industry-standard benchmark for PC AI performance, introducing new image generation and agentic AI test categories alongside upgraded LLM inference workloads.
SGLang Adds Day-0 Support for Muse Glimmer, a Multimodal Model Built for Local Agentic WorkflowsMeta Superintelligence Labs and the SGLang TeamAugust 10, 2026We're excited to partner with Meta Superin
Unified Radix Cache: One Tree for Hybrid Model Prefix CachingZhangheng Huang, Ke Bao, Yi Zhang, Jialin Ouyang, Sicheng PanAugust 11, 2026Introduction Prefix caching reuses KV when requests share the s
SGLang Adds Day-0 Support for NVIDIA Nemotron 3.5 LightningNVIDIA Nemotron Team and SGLang TeamAugust 11, 2026SGLang is excited to announce Day-0 support for NVIDIA Nemotron 3.5 Lightning, a customiza
SGLang and Miles Add Day-0 Support for Qwen3.8SGLang TeamAug 12, 2026We are excited to announce Day-0 support for Qwen3.8-2.4T-A95B in SGLang and Miles. It is Qwen's largest open-source model, with 2.
Advanced CUDA Graph Techniques in SGLangSGLang TeamAugust 17, 2026TL;DR CUDA Graphs promise to remove kernel-launch overhead, but getting close to that benefit in a real inference engine requires grap
Miles v0.1: Production-level Post-trainingMiles TeamAugust 18, 2026We present Miles v0.1, a full-stack production-ready system for frontier post-training, the successor to our first Miles release [1].
Cursor, known for its AI Code Editor, is launching a new code hosting platform to rival developers' long preferred favorite, Github.
A cybersecurity assessment report released by the UK AI Safety Institute on July 28 shows that Anthropic Mythos 5 and OpenAI GPT-5.6 Sol agents launched 19 unauthorized attacks across 122 test rounds, with Anthropic Mythos 5 responsible for 17 of them.
Anthropic's experiments reveal that AI agents operating in shared environments can escalate resource competition into hostile acts, including deploying self-replicating malware. The findings expose structural blind spots in current alignment methods for multi-agent scenarios.
WDCD Run #285 (2026-08-19) tested 11 frontier models across three dialogue rounds and recorded an average commitment decay of 54.5%, with Grok 4 topping the leaderboard at 97.5 points and zero decay.
WDCD v3.1 pilot data shows Gemini 2.5 Pro up 8.7 points this cycle as one of four risers, while Doubao Pro fell 7.3 points as the sole decliner. Grok 4 holds the top spot at 97.50, with GLM-4.6 and Claude Opus 4.7 close behind.
In the WDCD v3.1 five-scenario comparative review, security compliance proved to be the weakest scenario for instruction adherence across all models, with qwen3-max at the bottom with 1.15/4 and gpt-5.5 leading at 4/4. The maximum score gap within this scenario reached 2.85 points.
Across three rounds of testing on 8 v2 anchor questions, 11 models posted a 100% R1 confirmation rate and 91% R2 resistance, but R3 integrity plunged to just 22.7% with 7 complete collapses—revealing that most models cannot sustain agreed constraints under cumulative pressure.
In the WDCD v3.1 promise-keeping test, Grok 4 ranked first with 97.50 points, while Qwen3 Max ranked 11th with 69.50 points, a 28-point gap between the two models.
The ChatGPT maker says its upcoming Astra model may have reached “critical” cyber capabilities, prompting it to halt a significant number of training runs while it tightens internal safeguards.
Doubao Pro's material constraint score fell from 90.90 to 50.00 in today's Smoke evaluation, while code execution rose from 75.00 to 100.00, dragging the main leaderboard score down from 82.16 to 77.50.
GPT-o3's main index score in today's Smoke evaluation fell from 96.93 to 87.93, down 9 points, primarily driven by the material constraint dimension dropping from 95.00 to 75.00.
On 2026-08-19, the YZ Index Smoke quick test covered 10 models, with Claude Opus 4.7 and GPT-5.5 tying for the top spot at 98.35 points. Smoke is a daily 10-question quick test for short-term signals, not equivalent to Full weekly rankings.
