Granite 4.0 H Micro

Availability: This row is IBM's HYBRID 3B (4 attention / 36 Mamba2 layers), not the dense 3B Granite 4.0 Micro that IBM ships alongside it. The two are separate models on IBM's own model cards, and only the hybrid has a per-token price: Cloudflare Workers AI serves @cf/ibm-granite/granite-4.0-h-micro, while Artificial Analysis states the dense Granite 4.0 Micro "is not currently available through any API providers we benchmark". Until 2026-08-14 we published this price under the dense model's name and benchmarks; OpenRouter titles the same granite-4.0-h-micro slug "IBM: Granite 4.0 Micro", which is how the two were conflated. This is Cloudflare's price for hosting IBM's model, not a price from IBM: IBM's watsonx.ai rate card lists granite-4h-micro as "Not available" for per-million-token pay-as-you-go while pricing its siblings on the same table (granite-4h-small is USD 0.0636 in / USD 0.265 out), and sells it only as hourly on-demand GPU hosting, which is not comparable per token. The output rate is $0.112 to three decimals on Cloudflare's Workers AI pricing table; the model page rounds it to $0.11.