NeuralStackCohere announced Command R+ 2 on July 12 with a native 5 million token context window. The model achieves 92 percent on MMLU benchmarks and supports 23 languages at production quality. Enterprise customers can upload entire contract libraries for single-pass reasoning.
Training completed on July 5 using 18,000 H100 GPUs over six weeks. Inference pricing starts at $3 per million input tokens for context under 1 million. Early access opened to Cohere's 1,200 existing enterprise accounts.
Background includes Cohere's focus on retrieval-augmented generation since its 2021 founding. Command R+ 1 released in March 2025 introduced 1 million context. The new version removes the need for chunking in most legal workflows.
Why this matters
Command R+ 2 reduces reliance on vector databases for document-heavy industries. Law firms report cutting review time from days to hours on July 13 tests. This shifts competitive pressure toward models that handle full archives natively.
Financial institutions gain real-time compliance checks across years of filings. Developers can now build agents that maintain state across multi-year conversation histories. The release accelerates enterprise adoption of long-context AI beyond research labs.
Future updates will add native video and audio context by Q4 2026. Cohere plans to open a public leaderboard for 5M-context tasks next month.