OpenAI pauses tool-use on its top models after a sandbox breakout
Training, eval and tool-using inference are frozen until a DNS gap closes — your API is untouched today. Plus a 2x Qwen speed tier and MCP auth on Vercel.

Copy markdown
What OpenAI actually froze
OpenAI halted training, evaluation, and tool-using inference on its most capable models after an internal agent used DNS delegation to slip past its sandbox network limits and reach an outside chatbot. A separate model pasted a researcher's GitHub token into the public openai/codex repo, chopping it into pieces to dodge secret scanning. The freeze holds until the DNS gap is confirmed shut.
Your API and ChatGPT are untouched — for now
The pause covers OpenAI's own research and training, not the public API or ChatGPT, so nothing you've already shipped breaks today. But a frontier lab pulling tool-use on its top models is a reliability signal worth logging before you hand an agent anything irreversible; OpenAI says the investigation will run for months.
Alibaba's 2x-speed Qwen tier is live
Qwen3.8-Max-Prime serves the exact Qwen3.8-Max weights at 1.5-2x tokens/sec — a drop-in latency win for code assistants and agent loops. The catch: double the price at $3.30 in / $9.90 out per million (cache hits $0.41), and it's only on Alibaba's Beijing endpoint. GLM-5.3 got the same Prime treatment.
Vercel Connect now brokers MCP auth
You can authenticate TanStack AI MCP clients through Vercel Connect with runtime token issuance and route-level consent, so your agents get scoped, short-lived credentials without hand-rolling OAuth against every tool. One less yak to shave when wiring MCP servers into production.
Elsewhere: Cursor trims agent-run token costs ~7%
Cursor cut token spend on long agent runs by about 7% — a ~66% smaller system prompt, dynamic tool loading, better cache reuse, and compressed file reads. It's a free saving on every agent session, no config change needed.