Opus 4.8 at the same price as 4.7, with longer independent runs and better self-checking. Small version bump, big deal if you live inside agent loops all day.
Qwen3.7-Max is out. The headline stat for me: 35 hours on a kernel task, 1,000+ tool calls, no babysitting. Agent runway is starting to matter as much as context length.
Uber exhausted its entire 2026 AI budget on Claude Code and Cursor within four months.
Monthly API costs per engineer varied from $500 to $2,000 as usage surged. Now, 95% of Uber engineers utilize AI tools monthly, with 70% of committed code generated by AI.
GPT-Realtime-2 in the API — voice agents that can reason mid-conversation instead of just transcribing and guessing. That gap has felt obvious for a while.
ChatGPT add-in for Excel and Google Sheets. Forecasts and reconciliations via plain language will save hours. Audit trails are about to matter more.