Anthropic's models rank among the best when it comes to writing code. But they lag behind in voice and chat assistants for customer support. This is according to data from NiCE, which provides customer service software and counts Toyota, Allianz, Nestlé, and Frontier Airlines among its clients.
Speed is crucial
The problem lies in how long it takes the model to start responding after receiving a request. And for voice assistants, that is an issue. In customer support, every millisecond counts.
According to NiCE, a voice or text bot can remain silent for no more than roughly two and a half seconds. Any longer and it starts to feel unnatural to the customer. The market for AI-powered customer experience software is estimated at around $15 billion and is growing rapidly. If NiCE deployed Anthropic's models on its customer service line, the result would reportedly be terrible. The bot would sound awful because the models are simply too slow.
Google and OpenAI
NiCE recently compared data from its customers. Anthropic's Opus 4.7 model took more than four seconds to produce its first token. Google's Gemini 3 Flash needed only about one second. It also processed roughly four times as many tokens per second.
Anthropic's model costs 18.5 times more than Gemini 3 Flash. Anthropic does offer the faster Opus 4.6 Flash, which produces its first token more quickly than Gemini 3 Flash, but it costs nearly one hundred times as much as Google's model. The cheaper Sonnet and Haiku series are slower and slightly more expensive than Gemini 3 Flash.
Small, fast models from OpenAI, such as 5.4 mini and 4.1 mini, are also gaining popularity among NiCE customers. They are inexpensive. And at the beginning of the month, OpenAI launched a new audio model for ChatGPT that it does not yet offer through its developer interface. Even so, the model can do something extra. It processes the speaker's voice continuously, allowing it to react immediately when someone interrupts it. Much like in a conversation between people.
Where Claude still has the edge
Anthropic still retains one major advantage: accuracy. According to NiCE, the Opus 4.7 and 4.6 models make roughly half as many errors as Gemini 3 Flash. And that matters when customers expect first-class service.
An Anthropic spokesperson responded by saying that the company offers a full range of models because different tasks have different requirements. While models such as Fable and Opus are built for demanding reasoning that takes minutes to days, the Sonnet and Haiku series combine high intelligence with the speed needed for high-volume workloads. The quality of the finished work is reportedly consistent across the models.
Diversity is key
Atlassian has also expressed skepticism about having everything run through a single provider. The company positions itself as a conductor that helps customers switch between different models and coding assistants during software development.
Last week, Atlassian added a range of new features to Jira, its software project management application. Teams can now assign tasks, such as designing a website login page, to different tools. The options include Anthropic's Claude Code, Microsoft's GitHub Copilot, and Jira's own coding assistant.
Atlassian also improved its AI assistant for Slack, which bears a striking resemblance to Anthropic's recently introduced Claude Tag. Team members can tag the Jira assistant in their Slack channels—that is, group chats—and have it analyze messages and turn them into work items. The agent is also expected to work in Microsoft Teams soon.
