Cohere Releases Command A+: Open Weights, 86% Non-Hallucination, First Model in 14 Months
Cohere has released Command A+, an open-weights language model scoring 37 on the Artificial Analysis Intelligence Index. The release puts Command A+ in the same tier as Claude 4.5 Haiku, NVIDIA Nemotron 3 Super, and Gemini 3.1 Flash-Lite — capable mid-range models positioned for enterprise deployment rather than frontier reasoning.
It is the company’s first significant model release in over 14 months, following the original Command A.
Where It Leads
Command A+‘s clearest differentiation is on knowledge reliability. It scores 86% on AA-Omniscience Non-Hallucination, approximately three percentage points ahead of the next-best model at this intelligence tier. The trade-off is intentional: its AA-Omniscience Accuracy is 9%, which means the model tends to decline when uncertain rather than confabulate. For legal, financial, and compliance deployments — Cohere’s core enterprise markets — that calibration is the right one.
On output speed, Command A+ produces approximately 281 tokens per second on Cohere’s API, faster than Claude 4.5 Haiku, GPT-5.4 nano, and Grok 4.3, while still trailing Gemini 3.1 Flash-Lite Preview at 304 tokens per second.
Where It Trails
Command A+ has measurable gaps on the hardest benchmarks:
| Benchmark | Command A+ | Tier Context |
|---|---|---|
| HLE | ~11% | Below Haiku class average |
| GPQA Diamond | 76% | Below Haiku class average |
| Terminal-Bench Hard | ~25% | Well below coding-optimised models |
| SciCode | ~38% | Below scientific model class |
| MMMU-Pro (vision) | 63% | Between Haiku (59%) and GPT-5.4 nano (65%) |
The Terminal-Bench Hard gap is the most commercially relevant. At 25%, Command A+ is not a coding agent deployment option, even for tasks within its intelligence tier. That limit places it squarely in content, retrieval, and document-processing workflows rather than developer tools.
The Open Weights Angle
Command A+ is released with open weights — a significant shift for Cohere, which has historically positioned its models as API-only, enterprise-licensed products. Opening the weights is a competitive response to an environment where Qwen3.6-35B, Mistral Medium 3.5, and Llama 4 Scout have raised the baseline expectations for what open-weights models should offer.
For Cohere’s enterprise customers, the open-weights release enables self-hosted deployment, which has become a hard requirement for regulated-industry customers in financial services and healthcare across Europe — markets that Cohere targeted directly through its acquisition of Aleph Alpha earlier this year.
Context: Cohere After Aleph Alpha
Cohere’s $20 billion acquisition of Aleph Alpha gave it direct access to European government and financial services contracts, a dedicated on-premises deployment stack, and 30 researchers specialised in domain-specific AI. Command A+ is the first new Cohere model to enter those markets as a self-hosted option. Whether the combined entity can hold that enterprise position against DeepSeek V4 — which offers a substantially higher intelligence index score under an MIT licence — will determine how durable this release is commercially.