AI Infrastructure Partnerships Accelerate
Key Questions
What new features are included in Opus 4.8?
Opus 4.8 introduces effort controls, dynamic workflows, a cheaper fast mode, and improvements in honesty and reduced deception. It outperforms GPT-5.5 on most benchmarks while trailing in agentic terminal coding tasks.
Which companies are partnering on AI infrastructure in Europe?
VAST Data has partnered with Mistral to build NVIDIA GB300 AI factories across Europe. The alliance focuses on accelerated computing for advanced AI workloads.
How have AI model prices changed recently?
DeepSeek V4 Pro received a 75% permanent price cut, while MiMo V2.5 Pro is priced at $1/$3 per million tokens. These reductions highlight the rapid decline in AI inference costs.
What is driving the shift from LLMs to SLMs?
The industry is moving toward right-sized models that balance performance and efficiency for specific tasks. This trend supports broader adoption in resource-constrained environments.
How is agentic coding adoption progressing?
Cursor has reached $1B in revenue amid growing agentic coding use, with China at 30% adoption versus 91% in the US. Interactive agents and durable sandboxes are emerging as key evaluation metrics.
What improvements does MobileGym provide for vision-language models?
MobileGym boosts Qwen3-VL performance by 12.8% through a verifiable simulation platform that achieves 95.1% sim-to-real transfer. It enables scalable research on mobile GUI agents.
Which new open-source models were released?
Command A+ was released as an open-source model, alongside GLM5.1-NVFP4 and Mistral Vibe agent. HF async RL weight sync was also introduced for efficient training.
What does the data temporality study focus on?
The study examines how data freshness affects model performance and proposes methods to improve temporal relevance in training datasets. It addresses challenges in maintaining up-to-date AI systems.
Opus 4.8 officially released with effort controls, dynamic workflows, cheaper fast mode, honesty/deception improvements; beats GPT-5.5 on most benchmarks but trails on agentic terminal coding. VAST Data partners with Mistral to build NVIDIA GB300 AI factories in Europe. DeepSeek V4 Pro 75% price cut, MiMo V2.5 Pro pricing at $1/$3 per M tokens. Mistral Vibe agent launch, HF async RL weight sync, GLM5.1-NVFP4, Command A+ open-source. Agentic coding shift: Cursor $1B revenue, China 30% vs US 91% adoption. LLMs vs SLMs shift toward right-sized AI. Interactive agents TTI and durable sandboxes emerge as key metrics. MobileGym boosts Qwen3-VL by 12.8% with 95.1% sim-to-real. Polar decoupled rollout for RL. MiniMax M3 sparse attention 15.6x speedup. SAM state-adaptive memory. Data temporality study improves freshness.