OpenAI made GPT-5.5 the default on every paid tier on Thursday. The pitch is explicit: give it a messy, multi-part task and trust it to keep going. The flagship model's unit of work is no longer a good answer. It is a completed task. Every piece of context you give the agent now has to be good enough to survive unattended execution.

OpenAI shipped Workspace Agents in research preview: cloud-resident agents that run without your browser open, shared across a team, plugged into Slack and Salesforce. The architectural call is that the enterprise AI unit is not a user's chat session but a team-owned agent with its own permissions, memory and analytics. Free until 6 May. Worth standing one up this week to feel how the handoffs actually work.

Anthropic published an engineering post admitting three changes between March and April that quietly degraded Claude Code and the Agent SDK: a reasoning-budget cut on 4 March, a cache-clearing bug on 26 March, a system-prompt edit on 16 April reverted four days later. They reset usage limits. The more interesting thing is not that it happened. It is that they showed their working. Notice how rare that still is.

On Wednesday, Anthropic quietly updated their pricing page to remove Claude Code from the $20 Pro tier. Reversed within hours. Willison's piece is not primarily outrage. It is a practitioner working through what a silent pricing change means for a serious teaching bet. The load-bearing sentence: 'Should I be taking a bet on Claude Code if they might 5x the minimum price?' Read it for the argument shape.

GPT-5.5 ships with 'keep going' as its operating instruction. Workspace Agents make team-owned agents the enterprise unit. Anthropic's postmortem shows three silent config changes cut effective capability across three months. The agent surface is where the product energy is. The trust plumbing is where the variance is. The best AI workflows in 2026 will be run by the people who treat the plumbing as a first-class object.