IBM: Granite 4.2 8B

IBM Granite 4.2 8B is now routable through OpenRouter via CoreWeave as the sole provider, released August 31, 2026. Listed pricing is $0.10 per 1M input tokens and $0.15 per 1M output tokens, with cache reads at $0.05/1M. The weighted-average effective price customers actually pay is $0.05893 input and $0.1497 output p The model supports three reasoning modes—full, low-effort, and non-thinking—and covers 12 languages: English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese. Observed performance on the CoreWeave route is 96 tokens/second throughput (P50 best) with 0.20s P50 latency. Q

The eesel AI review of Granite 4.2 explicitly names granite-4.2-8b as one of the three released checkpoints and describes the headline change versus Granite 4.0: the model now plans and self-corrects inside ... tags before answering, with reasoning toggleable per request rather than baked in. The 8B is positioned for g The review also details the reasoning-augmented tool-calling loop, in which the model reasons about which tool to invoke and why before producing the call, and notes the architectural shift from Granite 4.0's hybrid mixture-of-experts design to a fully dense transformer in 4.2. The piece is dated August 29, 2026, with

IBM released Granite 4.2 as a reasoning-focused update to its open model family, with the 8B variant (granite-4.2-8b) post-trained from the granite-4.1-8b-base. The eesel AI overview names granite-4.2-8b as an 8B dense decoder-only transformer with 40 layers and a 128K context window, distinguishing it from the 3B and The same post documents that Granite 4.2 introduces switchable native reasoning in a single model, with full thinking, low-effort, and non-thinking modes, and that all 4.2 checkpoints are dense rather than the hybrid MoE architecture used in Granite 4.0. Reasoning is exposed via ... tags, enabling the model to plan and

LLM-Stats' entry for IBM Granite 4.2 8B, published August 25, 2026, places the model at rank 198 on its composite LLM Stats Score of 19.8, with a blended price around $0.069 per million tokens. It reports a capability tier of B (average, top half) and per-category standings: Math 148 of 326, Long Context 91 of 119, Too Individual benchmark rows include AIME 2025 at 0.87 (rank 62), RULER 64k at 0.81 (rank 3), IFBench at 0.79 (rank 10), and HMMT25 at 0.78 (rank 14), with 14 benchmark scores listed in total. The cost-efficiency comparison frames Granite 4.2 8B's $0.069 blended price against models ranging from Gemma 4 E4B at $0.024 to G

BenchLM's profile for Granite 4.2 8B records the model as released August 25, 2026, open-weight, with 128K context and a reasoning capability. The aggregator assigns Granite 4.2 8B a composite score of 42/100, ranking it 174 of 232 tracked models, with its strongest eligible category being Instruction Following at 45. On per-axis percentiles, Granite 4.2 8B lands at the 63rd percentile for Instruction Following, 14th for Agentic, 23rd for Coding, and 13th for Knowledge, while Reasoning, Math, Multilingual, and Multimodal are flagged as not rank-eligible in this corpus. Speed is reported at 144 tok/s with a first-token latency of 14.