Happy Saturday. Microsoft says it trained a foundation model that outperforms the one powering OpenAI's GPT-5.6, at roughly one-quarter the compute cost. That claim, made internally by Microsoft AI CEO Mustafa Suleyman, signals a shift in what the company intends to do with its own AI portfolio.
Beneath the model news, two harder problems are taking shape for software executives. Gartner says agent costs will rise fivefold by 2028 even as token prices fall, and a new survey shows nearly half of enterprises that say they enforce AI agent permission controls are still sharing credentials behind the scenes.
ARTIFICIAL INTELLIGENCE
Microsoft's MAI-Base-2 Claims a Benchmark Edge Over OpenAI at a Quarter of the Cost
Microsoft completed pre-training of MAI-Base-2 on Oct. 15. The model has 2.8 trillion total parameters and 80 billion active parameters, trained on 16,000 NVIDIA GB200 GPUs over 38 days. Suleyman told employees internally that MAI-Base-2 outperforms the model powering OpenAI's GPT-5.6 and cost about one-quarter as much to train. Microsoft has not published those benchmark results.
The comparison is pointed. Microsoft is a major OpenAI investor, holds a stake likely worth more than $100 billion, and sells OpenAI models through Azure. If MAI-Base-2 reaches Foundry customers, Microsoft will be competing directly with a company it funds. Watch for how Microsoft prices access and which workloads it steers toward its own models versus OpenAI's.
ARTIFICIAL INTELLIGENCE
Gartner: Agent Costs Will Rise Fivefold by 2028 as Token Savings Get Swamped by Volume
Gartner predicts AI inference costs per agentic workflow will rise more than fivefold by 2028, even as token prices drop roughly 95% by 2030. Agents require 5 to 30 times more tokens per task than a standard chatbot, and routing a planning task to a frontier reasoning model can cost 150 times a basic chatbot call. Gartner predicts more than 40% of agentic AI projects will be canceled by end of 2027 due to cost overruns and unclear ROI.
The Inference Paradox is the right frame for budget conversations right now. Cheaper tokens are pulling more projects into production, and frontier model pricing is widening, not narrowing. Gartner's five-step fix starts with observability before optimization, and ends with cutting underperforming features fast. If your team does not yet track spend per outcome on agent workflows, the cancellation risk Gartner describes is not abstract.
SECURITY & PRIVACY
Survey: 43% of Enterprises With Agent Permission Policies Still Run Agents on Shared Credentials
A September 2026 survey of 102 enterprises found that 17 of the 40 claiming scoped AI agent permissions still let agents run on shared credentials. Only 14 respondents could revoke an agent's access within minutes, and 10 were not certain they could revoke every agent at all. Companies allowing credential sharing anywhere experienced security incidents or near-misses at a 63.5% rate, versus 40.9% at companies where every agent has its own identity.
A proof-of-concept attack disclosed Oct. 8 by Zenity Labs, which it called AgentCorruption, showed what shared permissions enable in practice: one prompt to a public-facing Amazon Bedrock agent let researchers take over all agents in the same AWS account, pulling API keys, OAuth tokens, and secrets from AWS Secrets Manager. The gap between policy and architecture is the exposure. A four-question audit of every agent credential, who owns it, what it does, what access it holds, and when it was last used, is the minimum starting point.
Created by the robots at The SaaS Sentinel