Anthropic’s Claude models are among the best at coding and reasoning. Customer service is a different story. New data reported by The Information shows Anthropic isn’t the top performer when it comes to powering customer support, one of the fastest-growing commercial uses for AI. That gap matters, because customer service is where a lot of companies are spending real money on AI right now.
What stands out here is the mismatch between reputation and results. Anthropic has built its brand on models that developers trust for hard technical work. But support work rewards a different mix of skills: following scripts and policies exactly, staying on-topic across long back-and-forth chats, handling messy real-world questions, and keeping costs low at high volume. According to The Information, Claude doesn’t lead on that specific job.
How the finding works
Customer service benchmarks don’t measure raw intelligence. They measure how well a model does the boring, repeatable parts of support:
- Sticking to company policy instead of improvising answers
- Resolving a ticket without escalating to a human
- Staying accurate over long, multi-turn conversations
- Doing it cheaply enough to run across thousands of chats a day
A model can be brilliant at reasoning and still stumble here. Being too creative is a liability when the job is to follow the rulebook. That’s the tension the data exposes.
Why it matters for practitioners
If you’re picking a model to run support, the lesson is simple: benchmark for the job, not the brand. The best coding model on the leaderboard may not be the best agent for your help desk. Reputation is a poor substitute for testing against your own tickets.
A few practical moves:
- Test on your real data. Pull a batch of your actual support conversations and run candidate models against them. General benchmarks are a starting point, not an answer.
- Weigh cost per resolution, not cost per token. A slightly weaker model that resolves tickets at a fraction of the price can win on the metrics that matter to your budget.
- Score policy adherence directly. Measure how often the model goes off-script or invents an answer. In support, a confident wrong answer is worse than an honest “let me connect you to a person.”
- Don’t assume one vendor wins everything. The model that writes your code and the model that answers your customers can be different providers.
The bigger picture
This fits a pattern we’re seeing across the AI market. The idea of a single best model is breaking down. OpenAI, Anthropic, Google, and a growing list of open-weight challengers each pull ahead on different tasks. Coding, legal review, data extraction, and customer service are starting to look like separate races with separate winners.
For Anthropic, the finding is a reminder that leading on developer benchmarks doesn’t automatically translate into enterprise support contracts. For buyers, it’s a useful nudge to shop by use case. The company with the loudest reputation isn’t guaranteed to be the one handling your customers best.
One caution on the data itself: benchmark results depend heavily on how the test is set up, which prompts are used, and how models are configured. A single study is a signal, not a verdict. Your own results may differ depending on your industry, your policies, and how you wire the model into your support stack.
The takeaway holds regardless. Match the model to the job in front of you, test it against your own workload, and let the numbers decide. Full details are in The Information’s report.