Cohere Launches Open Weights Model Command A More Than A Year Since The Command A Release

Cohere has released Command A+, an open-weights model that arrives more than a year after Command A. Artificial Analysis scores it 37 on its Intelligence Index, in line with Anthropic's Claude 4.5 Haiku and just above NVIDIA Nemotron 3 Super and Google's Gemini 3.1 Flash-Lite (artificialanalysis.ai).

Its clearest strength is on AA-Omniscience, Artificial Analysis's knowledge and hallucination evaluation. Command A+ answers only 9% of its questions correctly, which is low, but it also hallucinates the least of any model tested: it ranks first on the Non-Hallucination measure at 86%, about 3 percentage points ahead of the next-best model. Because the evaluation counts wrong answers against a model, that combination yields a headline score of -4, which Artificial Analysis calls fairly strong relative to other models. It describes Command A+ as a model that knows its limits, the same pattern it sees in Claude 4.5 Haiku.

On Cohere's API it produced about 281 output tokens per second, faster than GPT-5.4 nano, Claude 4.5 Haiku and Grok 4.3 but slower than Gemini 3.1 Flash-Lite Preview at 304. Artificial Analysis says that gives it more intelligence and speed than gpt-oss-120b, though it sits behind the frontier of best speed-for-intelligence trade-offs that Gemini 3.5 Flash set the day before. The weak spots are scientific reasoning (about 11% on HLE, 76% on GPQA Diamond) and coding (about 25% on Terminal-Bench Hard, 38% on SciCode), where it trails its peer set. It also handles images, scoring 63% on the MMMU-Pro visual reasoning test, between Claude 4.5 Haiku at 59% and GPT-5.4 nano (xhigh) at 65%.