AI digest: Big model battles and infrastructure plays
NVIDIA pushes 4-bit training to new limits, Cursor matches frontier models at lower cost, and Anthropic makes strategic moves in tooling and enterprise security.
The week brought real advances in model efficiency alongside some predictable industry consolidation moves.
NVIDIA cracks 4-bit pretraining at massive scale
NVIDIA validated 4-bit pretraining on a 12B hybrid Mamba-Transformer across 10 trillion tokens, the longest documented run at this precision. Their NVFP4 approach matches FP8 baseline accuracy whilst cutting memory requirements dramatically. This matters because it makes large-scale training accessible to more teams, not just the hyperscalers. Source
Cursor’s Composer 2.5 matches frontier models for less
The coding assistant now performs on par with Opus 4.7 and GPT-5.5 on benchmarks whilst costing significantly less to run. Built on Kimi K2.5 and trained on 25x more synthetic tasks than the previous version. Smart positioning for developer tools where cost per query actually matters in daily use. Source
Anthropic acquires Stainless for SDK automation
The Claude maker bought the dev tools startup that automates SDK creation and maintenance for APIs. Stainless already serves OpenAI, Google, and Cloudflare, so this gives Anthropic immediate credibility in developer tooling. Makes sense as API complexity grows and companies need better ways to ship integrations. Source
Claude Mythos finds critical financial system flaws
Anthropic will brief global financial regulators on cyber vulnerabilities its new Claude Mythos Preview discovered in banking infrastructure. This positions Claude as more than a chatbot, moving into specialised security analysis where AI can process codebases at scale humans simply cannot. Clever enterprise positioning beyond the usual productivity use cases. Source