Proprietary on Best of AI

Command A2

Sun, 19 Jul 2026 00:00:00 +0000

Command A2 succeeds Command A as Cohere’s flagship model, keeping the same enterprise-first pitch: strong retrieval-augmented generation, tool use, and multilingual support over chasing chatbot benchmarks. It’s priced at $2.50 per million input tokens and $10 per million output tokens, and it’s reachable through Cohere’s own API as well as AWS Bedrock, Azure AI Foundry, and Oracle Cloud, the same multi-cloud distribution Cohere has relied on since Command R.

Cohere continues to sell mainly to enterprises building internal search and document-QA systems rather than consumer-facing products, and Command A2’s improvements are aimed squarely at that audience: better grounding on retrieved documents and more reliable citation of sources, both areas where the company has staked its reputation against larger labs with flashier general-purpose models.

Mistral Large 4

Sun, 19 Jul 2026 00:00:00 +0000

Mistral Large 4 replaces Large 3 as Mistral AI’s proprietary flagship, roughly eight months after that model shipped in December 2025. It’s a 190 billion parameter dense model with a 256,000 token context window, priced at $2 per million input tokens and $8 per million output tokens through Mistral’s own La Plateforme as well as Azure AI Foundry and AWS Bedrock. Mistral says the gains over Large 3 are concentrated in coding and multi-step reasoning, an area where the company has been trying to close the gap with GPT-5.6 and Claude Sonnet 5 rather than compete on raw scale.

Qwen4-Max

Sun, 19 Jul 2026 00:00:00 +0000

Qwen4-Max is Alibaba’s top-tier proprietary model, sitting above the open-weight Qwen3 family the company keeps releasing on Hugging Face. Unlike those, Qwen4-Max is closed and only reachable through Alibaba Cloud’s API, Qwen Chat, or partner platforms, priced at $1.20 per million input tokens and $4.80 per million output tokens. It takes text and image input, carries a 1 million token context window, and Alibaba has been positioning it directly against GPT-5.6 and Gemini 3.5 Pro in its own benchmark comparisons.