AI news this week: Haiku 5.5's price cut, GPT-6.1 Terra, and a policy shift you'll feel
The late-October wave was about cheaper models and computer-use agents. The early-November story is who gets oversight.
The last two weeks of October 2026 were busy enough that the headlines blurred. Underneath them, a pattern: the expensive flagships stayed put while the cheap models got radically better, and the guardrails moved from suggestion to rule.
Anthropic shipped Claude Haiku 5.5 on October 7 at roughly $0.10 per million input tokens, about 75 percent cheaper than the prior Haiku. For high-volume, low-stakes tasks, that changes the math more than any benchmark.
OpenAI released GPT-6.1 Terra on October 26, a model built for computer use and coding at $1 per million input and $5 output. ChatGPT itself moved to GPT-6 with an 'Intelligent UI' that bakes charts, buttons, and forms into answers. GPT-5.5 was retired from the consumer app on October 14.
On the open side, Mistral Large 4 entered preview as a 1-trillion-parameter open-weight model, Moonshot's Kimi K3.1 Pro promised weights by early November, and DeepSeek dropped V4.1-Flash with MIT-licensed weights and a million-token context. Builders who want to self-host finally have real options.
The part that will actually touch users: Microsoft made Execution Containers generally available on Windows 11, sandboxes that cap what an agent can reach on your files and network. And Anthropic set a usage policy, effective November 12, that requires human oversight for equipment that could cause injury. The agent economy is getting a leash, slowly and sensibly.
Frequently asked questions:
Q: Which model is cheapest right now?
A: Claude Haiku 5.5 and GPT-6 Luna sit at the bottom of list pricing for short prompts.
Q: Should I care about Execution Containers?
A: If you let agents touch your machine, yes. It is the difference between trust and hope.
Q: Are open weights real this time?
A: Mistral, Kimi, and DeepSeek all published or promised weights in late October. Self-hosting is back on the table.
