IBM Granite 4.2 8B lands on OpenRouter: reasoning at $0.10/$0.15
A dense open-reasoning model with a 131K window, hosted so you skip the GPU — plus Apache-2.0 weights if you'd rather run the 3B, 8B, or 30B yourself.
Copy markdown
$0.10/$0.15 per million, no GPU to rent
Granite 4.2 8B went live on OpenRouter on Aug 31 at $0.10 input / $0.15 output per million tokens, with a 131K-token context and matching max output. CoreWeave is the sole host so far at ~0.13s latency — cheap enough to leave a reasoning loop running overnight.
A dense 8B trained to plan before it acts
It's an open-reasoning model with native step-by-step thinking, and the 8B specifically got an agentic-RL training phase aimed at software engineering and terminal work. That makes it a plausible cheap backend for coding agents and tool-use loops, not just chat.
Apache 2.0: run the 3B, 8B, or 30B yourself
The full 4.2 family (3B / 8B / 30B) ships under Apache 2.0 on Hugging Face, so if the hosted route isn't private enough you can pull the weights. The 3B fits a laptop; the 30B suits a dedicated local reasoning box.