OpenAI's latest flagship model ships with integrated reasoning chains and live browsing — the first time both capabilities have launched simultaneously. API access is live now, with consumer rollout beginning this week.

The latest Opus model sets a new benchmark for long-context reasoning, processing entire codebases and legal documents in a single pass. Anthropic says it outperforms all current models on multi-document analysis tasks by a significant margin.

Detailed compliance guidance for high-risk AI system providers is now live, with requirements taking effect from August. Companies operating in Europe need to act quickly — the document names specific product categories requiring independent audit.

In a four-hour conversation, DeepMind's CEO makes the case that current scaling trajectories are being underestimated, and outlines what the next two years of AI development will look like. One of the most substantive interviews of the year.

A peer-reviewed study shows a 94% accuracy rate — 11 points above specialist average — with significantly lower false positive rates across all cohorts. Published in Nature.

The Paris-based lab is expanding compute infrastructure and research hiring through 2026, backed by General Catalyst and Andreessen Horowitz. Europe's AI race is real.

Nikhil Bier's thread breaks down the architectural differences in the new Grok release, including significant improvements to long-context handling and code generation benchmarks.

The federal government's framework is voluntary for now but signals the shape of what's coming. Relevant if you operate in or sell into Australia — mandatory requirements are expected within 18 months.

A major update to AlphaFold extends its predictions beyond protein structures to drug binding sites. Pharmaceutical researchers say this could halve early-stage drug discovery timelines for certain disease classes.

In a long thread following the GPT-5 launch, Altman outlines his view on societal implications — less philosophical than usual, more concrete. Worth reading in full.

The former OpenAI and Tesla AI director has published a comprehensive video series walking through building large language models from first principles. Already has 500k views in 24 hours.

The New Zealand government has formed a dedicated AI task force to develop a national framework, the first formal step toward AI regulation in the country. Industry consultation opens next month.

ByteDance's Seed-Coder model is available under a permissive licence and reportedly matches GPT-4 on standard coding benchmarks. Independent evaluations are underway.

An engineering lead at a mid-size fintech shares what actually broke — prompt injection attempts, cost overruns, and hallucination patterns in financial summarisation tasks. The most grounded AI-in-production write-up published this month.