This week the frontier was about contestability — who can defend their model's IP, who controls the consumer market, and how quickly a model generation can be displaced. Anthropic took an alleged model-theft campaign to the US Senate, fresh venture money rotated toward robots that learn from video games and drugs designed by AI, and the once-unassailable ChatGPT quietly lost its last GPT-4-era model and, by some measures, its majority of the assistant market. The through-line: nothing at the frontier is settled — not the IP, not the market share, not the reigning model.
Anthropic takes a distillation fight to the Senate
Anthropic told US senators that Alibaba's Qwen lab ran what it called the largest known distillation attack on its models, using roughly 25,000 fake accounts and 28.8 million exchanges between late April and early June to extract Claude's capabilities. Senators are now drafting an amendment to sanction or blacklist those responsible. Beyond the immediate accusation, the episode reopens a harder question: is distillation even theft? However that line is drawn, model access — not just chips — looks set to become an export-control battleground between the US and China.
Robots that learn from video games
General Intuition raised a $320M Series A at a $2.3B valuation on the bet that gameplay is the best training data for the physical world. The company is building spatial, embodied frontier models trained on billions of game clips drawn from some 17 million players via the Medal platform, arguing that button-press-timed gameplay teaches agents how to act in space better than scarce teleoperated robot data. Khosla Ventures led the round, with Jeff Bezos and Eric Schmidt among the backers — a cap table that signals how seriously the embodied-AI data thesis is now being taken.
AI drug discovery starts to show its receipts
At the BIO 2026 convention, AI-discovered drugs moved from promise toward proof, with the anchor signal a collaboration worth up to $2.5B between Insilico Medicine and SK Biopharmaceuticals on neuroimmune and central-nervous-system disorders. The deal is concrete evidence that AI drug discovery is converting into pipeline value rather than demos — and it runs against a tense geopolitical backdrop, with the BIOSECURE Act shadowing how US and Chinese biotech firms partner.
The quiet end of the GPT-4 era
With little fanfare, OpenAI removed GPT-4.5 from ChatGPT, the last model of the GPT-4 family in the product, with conversations now falling back to GPT-5.5. The model line that defined the modern chatbot era is gone from the flagship product (API access remains a separate track), and the retirement is as much symbolic as practical: it marks how sharply the frontier now compresses a model's shelf life, and how the cadence of retiring and re-anchoring flagships has itself become a competitive variable.
ChatGPT's grip on the assistant market slips
For the first time, ChatGPT has fallen below half of global AI-assistant usage. According to widely-cited usage trackers, its share slipped to roughly 46.4% as Gemini (around 27.7%) and Claude (around 10.3%) gained ground — figures drawn from usage-tracker roundups rather than an official OpenAI release, so best read as directional. The milestone, if it holds, reframes the consumer AI layer as a genuine three-way contest in which distribution and defaults may matter as much as raw model quality.
The money this week followed two currents. The first is contestability of the consumer layer: a sub-50% ChatGPT and the retirement of the last GPT-4-era model together suggest the assistant market is now genuinely up for grabs, with distribution and defaults rising in importance. The second is rotation: fresh frontier dollars are landing less on bigger chatbots and more on embodied AI and AI biotech, where General Intuition's $2.3B valuation and Insilico's up-to-$2.5B deal mark the new surfaces. Watch the Senate's language on distillation, whether gameplay-derived data proves out for robotics, and whether Gemini and Claude can sustain their share gains.