This Week in AI: Agents Go to Work, Models Get Cheaper

The story this week wasn't a single model. It was agents showing up inside the tools people already use. Google opened a persistent Gemini agent for work on October 8 that runs in the cloud, keeps context across devices, and picks up tasks after your laptop closes. Microsoft answered with Autopilot, a 'digital coworker' with its own identity and configurable permissions, and natural-language app building inside Copilot.

On the model side, Anthropic shipped Claude Haiku 5.5 on October 7 at $0.10 per million input tokens for prompts under 100,000 tokens, roughly 75 percent cheaper than the previous Haiku. That same day it halved Sonnet 5.5 cache-read pricing to $0.10 per million, a direct nod to GPT-6 Luna's economy tier. Google's Nano Banana 2.1 image model and Mistral's Large 4 preview, a 1.05-trillion-parameter mixture of experts, both landed October 6.

Perplexity quietly became the cheapest way to make a decision: pplx-decider-v1-27b, out since October 1, costs $0.04 per million input tokens with free output. For the thousand small routing calls a workflow makes, that undercuts the big general models by a wide margin.

The thread connecting it all: labs are racing on price and autonomy, not just raw capability. For anyone picking a model this month, the right move is to test the cheap tier against whatever you run today before assuming you need the flagship.

None of this is a reason to rewrite your stack. It's a reason to re-benchmark the boring, repeated calls, because that's where the savings actually show up.