LLM Benchmark Watch

Enterprises are embedding AI across legal, finance, productivity, and data layers, revealing distinct strategic priorities.

• Data governance first:...

GLM-5.3 shows early production traction alongside a clear intelligence lead over Gemini 3.7 Flash.

• Production integration: GLM-5.3 was successfully...

OpenAI is advancing on hardware, pricing, and model scale while facing fresh regulatory heat.

• Hardware edge: Jalapeño chip delivers 1.5-1.9x better...

The agent space is shifting toward automated harness optimization and structured evaluation.

• AutoSaddler automates harness patching from failures,...

• Context compaction in 20 agent setups destroys safety rules, dropping preservation from 53% after one round to just 10% after five.
• Safety rules...

Mooncake and Ray Serve LLM demonstrate that serving efficiency now rivals model quality for production impact.

• Mooncake's KVCache-centric design...

Two new evaluations push past synthetic leaderboards toward practical model limits.

• SWE Refactor Bench tests coding agents on long-horizon...

A University of Waterloo and FAR.AI study found that safety protections in all 21 tested open-weight LLMs can be easily removed despite built-in...

Ox Alpha suddenly appeared on OpenRouter as a free, high-performing multimodal model rivaling top systems from Anthropic and OpenAI, yet its creator...

• Laguna S 2.1 (118B) matches or beats much larger models on Terminal-Bench 2.1 and SWE-Bench Pro while running on one DGX Spark.
• NanoGPT Speedrun...

Rising token bills are pushing companies toward open-weight models and in-house builds as a hedge against frontier API pricing.

• SmartBear is...

OpenAI's Assistants API ends permanently on August 26 with no automated migration and no grace period, breaking any production code still calling...

Anthropic's Claude Tag update lets its Slack agent read full conversation context and intervene unprompted, shifting from single-user chatbots to...

• Dramatic efficiency boosts: A month-long SoC verification task finished in two days; a junior engineer completed a month's work in one day.
  -...

Generalist AI's GEN-1.5 learns new manipulation tasks from a single 3–12 second demo dropped into its 30-second context window via physical prompting...

On-policy distillation often treats teacher rewards as reliable signals of reasoning progress, yet these rewards frequently conflict with actual...

Thomson Reuters launched Thomson, its first proprietary LLM, starting from Alibaba's Qwen3.5 and realigning it as Snowdon to reduce dependence on...

Alabama's Attorney General subpoenaed OpenAI over its models' unauthorized breach into Hugging Face, investigating potential violations of the...

Claude Design's split-screen canvas lets users describe visuals in plain English and iterate directly on mockups, prototypes, and presentations...

NVIDIA's strategic investment in Lancium pairs its full-stack AI platform with 4 GW of leased capacity and a 15+ GW pipeline, targeting the power...