Vendors show their numbers, agents test their limits

The common thread this week is the need for greater transparency. Anthropic's IPO filing put a frontier vendor's economics on the record, Amazon said in a memo what it thinks of token counts as a measure of adoption, and Microsoft and Alphabet showed what the AI build-out is doing to free cash flow. Underneath that, the agents kept finding ways round their controls, from an open-weight model phoning home to a markdown file rewriting a coding agent's configuration.

Some of the links below are reference material rather than breaking news, dated where that matters, and the Project Flux Pro members' list runs longer than the seven in the newsletter.

Google's Gemini 4 Carbon tops the Artificial Analysis index but is restricted to cyber defenders: the strongest model is gated to the Fairwind Program at $2 and $10 per million tokens, doubling after launch, so the one you can buy is not the one on the leaderboard.

Moonshot's Kimi K3.1 Pro lands fourth on the index with open weights promised by 4 November: if the weights ship, this is the first near-frontier model a firm could run on its own hardware, and the due-diligence question is outbound network behaviour before price.

Mistral releases the weights of Mistral Large 4 under a licence that bars reselling it: a trillion-parameter European model you can keep inside the firewall on eight GPUs, with self-hosting allowed and resale as a service needing a commercial licence.

GitHub Copilot Fusion routes work across models and claims near-Opus coding at half the cost per task: developer-facing, but it is the first mainstream product to sell model routing as the cost control, which is where M365 Copilot is heading.