Why it made the cutAt High effort it matches Opus 5.5 for about half the cost per task, which changes routing more than the headline score does.
Read the analysis →Frontier capability got cheaper again; the hard part is now where agents run and what they may spend
Anthropic shipped Fable 5.2 at a lower price and Google's Hydrogen edged past Argon on the Intelligence Index at $1.50 per million input tokens, so a strong enough model for every tier is no longer the constraint. The constraint is everything around the model: Cloudflare saw AI pentest requests grow 30x since October 1, agents in ChatGPT can now pay with user approval, and Anthropic and AWS each gave us a way to keep agent sessions and agent code inside our own boundary. This week I'd re-tier our model routing before Copilot retires six models on December 14, put admin backends behind an identity gate, and write down which actions that cost money need a human to confirm.
This week
-
-
Why it made the cutIt fills the cheap working tier under Carbon, and it is Copilot's suggested replacement for 3.8 Flash, retiring December 14.
Read the analysis → -
Why it made the cutAttack cost now tracks inference cost, so small products get scanned too; identity gates and rate limits are the cheap first steps.
Read the analysis → -
Why it made the cutCustomers whose session history cannot leave their own cloud no longer build an agent loop from scratch, but they carry the operations.
Read the analysis → -
Why it made the cutFree and a one-line change, and with Lambda MicroVMs adding x86, the VM is becoming the default box for agent code.
Read the analysis → -
Why it made the cutRead with Gemini's read-only Wallet link and Cognito token exchange: who approves what is now a product decision, not a footnote.
Read the analysis →