AI digest: Big moves and bigger bills
Major model releases from Google and NVIDIA, Anthropic hits profitability, and OpenAI claims to solve 80-year-old maths problems.
The AI world keeps moving fast. This week brought performance breakthroughs, business milestones, and some seriously expensive compute deals.
Google’s Gemini 3.5 Flash beats flagship models at half the cost
Google dropped Gemini 3.5 Flash at I/O 2026, and it’s genuinely impressive. The model beats Google’s own flagship on coding and agentic benchmarks while running four times faster and costing 50% less. This feels like the first time we’ve seen clear efficiency gains without sacrificing capability, which matters more than the flashy demos.
NVIDIA’s Nemotron-Labs-Diffusion processes 6x more tokens per forward pass
NVIDIA released Nemotron-Labs-Diffusion, a model family that supports three different decoding modes in one architecture. The standout feature is 6x more tokens per forward pass compared to Qwen3-8B, which could genuinely change inference economics. We’re seeing real innovation in model architecture rather than just scaling up parameters.
Anthropic hits profitability while paying xAI £1.25 billion monthly
Anthropic told investors it’s about to have its first profitable quarter, with revenue doubling to £8.7 billion in Q2. Meanwhile, they’re paying xAI £1.25 billion per month for compute access. The fact that Anthropic needs external compute despite profitability shows just how constrained GPU supply still is.
OpenAI claims to solve 80-year-old geometry problem
OpenAI says its reasoning model disproved a geometry conjecture unsolved since 1946. This time the mathematicians who exposed their last embarrassing claim are backing it up. If genuine, it’s a proper breakthrough in AI mathematical reasoning, not just benchmark gaming.