Gemini 4 Carbon Tops AI Index; Anthropic Sits on $131B; OpenAI Eyes Google Chips
The compute supply chain is diversifying as OpenAI turns to Google for chips, and Anthropic's $131B cash pile means it can fund aggressive training runs without outside money. Developers building on any of these platforms should expect shifting pricing and availability as vendors race to secure capacity.
Google DeepMind's Gemini 4 Carbon scored 59 on the Artificial Analysis Intelligence Index v5.0, taking the top spot and also leading Text Arena with 1528 points. The model claims strength in cybersecurity and long-horizon tasks, and it follows Argon by just two weeks, showing a rapid release cadence. OpenAI announced three unverified proofs in mathematics, computer science, and physics, but refused to publish them, drawing criticism from the academic community.
On the hardware side, Bloomberg reports OpenAI will buy AI chips from Google, pulling a direct competitor's TPUs into its compute plans. Apple launched its first touchscreen MacBook Pro with M5 Pro/Max, starting at 22,999 RMB. Microsoft's multi-agent diagnostic system MAI-DxO hit ~83% accuracy on 304 difficult clinical cases, versus ~36% for doctors who could consult materials, according to a paper in Science.
Anthropic's IPO filing shows $131.7 billion in cash and short-term investments, more than Amazon or Microsoft. The company's first-half interest income alone covered its operating loss. Separately, the DOJ is asking questions about Nvidia's $12.93 billion acquisition of Hugging Face. Google Cloud revenue jumped 103% year over year to $30.7 billion, with model APIs processing nearly 32 billion tokens per minute.
Anthropic's cash position—$131.7B against a $693M operating loss—means it can run training and inference for years without having to compromise with cloud partners or investors. That changes how it negotiates compute deals and model pricing.
OpenAI buying from Google is a direct admission that Nvidia supply alone is not enough. The fact that one of its chip leads previously worked on TPUs suggests the engineering path is already known.
The dual-use AI security story—a Claude model building five attack chains while a startup breached Anthropic with agents—shows that the same technology enabling powerful coding assistants is also lowering the barrier to offensive security work.
Gemini 4 Carbon shipping two weeks after Argon indicates Google DeepMind is now on a monthly or faster cadence for frontier releases, which pressures OpenAI and others to accelerate their own timelines.