Skip to content
Join our Newsletter

New Cohere AI model's performance rivals latest versions of DeepSeek, ChatGPT

TORONTO — Cohere Inc. says it has released a new enterprise generative artificial intelligence model that is "on par or better than" the latest technology underpinning ChatGPT and DeepSeek.
64b90e6ee8a44b9ea9f3243d6408134c70e40fe008de0b71a2b0fb25b18c6579
Nick Frosst, co-founder of Cohere, is shown at the AI company's offices in Toronto on Monday, Nov. 27, 2023. THE CANADIAN PRESS/Chris Young

TORONTO — Cohere Inc. says it has released a new enterprise generative artificial intelligence model that is "on par or better than" the latest technology underpinning ChatGPT and DeepSeek.

The Toronto-based firm revealed the new product — Command A — on Thursday, advertising it as offering maximum performance with minimal hardware costs.

Command A is well suited for tasks like coding and answering technical or human resources questions in a requested language. Such agentic tasks tend to involve sophisticated reasoning or many complex steps and are increasingly being touted by the industry as the latest promising frontier for AI.

DeepSeek, the Chinese-made AI assistant whose debut earlier this year came with claims of better performance and lower costs than OpenAI's rival product ChatGPT, has pushed companies that have spent millions in the race to innovate with such technology to look for ways to outdo their competitors while using smaller budgets.

Cohere said it spent less than US$30 million to train Command A.

"We’re proud that Command A outperforms other leading models on a wide range of metrics relevant to our enterprise customers — and we did it using significantly less compute resources," Cohere co-founder Nick Frosst said in an email to The Canadian Press.

That efficiency can be seen in the number of graphics processing units — expensive chips needed to power AI models — Cohere's model can run on. Command A is deployable on just two GPUs, compared with other models that can require as many as 32, Cohere said.

DeepSeek has claimed building its assistant took two months, cost about US$6 million and used some of Nvidia’s less-advanced H800 semiconductors rather than the higher computing power needed by other AI models.

The company did not immediately respond to request for comments about how it thinks its product stacks up against Command A or doubts some have raised about DeepSeek's figures because of the much higher sums being spent by competitors and reports suggesting the company had access to tens of thousands more GPUs than it said it used.

When Cohere tested its model against GPT-4o and DeepSeek-V3 with coding and answering technical questions, it found Command A was in the same league and in some cases, ahead of GPT-4o and DeepSeek-V3.

ChatGPT-maker OpenAI did not immediately respond to a request for comment about the claims.

Frosst said Command A is particularly good at helping users securely get responses to questions based on internal company data and tools.

When combined with Cohere's agentic workplace platform North, he said it can also help users work more productively on tasks like analyzing long reports or summarizing data from sources like emails

"The model is also highly multilingual, so it’s ideal for global enterprises with workforces around the world," he said.

"Because it can run on just two GPUs, Command A is also great for private deployments, so customers in regulated industries can use it to build applications and power agents."

It currently supports 32 languages, which Cohere said makes it "much better" than GPT-4o or DeepSeek-V3 at consistently answering with content in the requested language.

This report by The Canadian Press was first published March 13, 2025.

Tara Deschamps, The Canadian Press