Anthropic is charging enterprise customers more for Claude while delivering less of it.
The company is shifting Claude Enterprise from a flat fee of up to $200 per user per month to a model that adds per-token computing charges on top of a $20 per user base fee. Fredrik Filipsson, co-founder of Redress Compliance, told Gizmodo the change could triple costs for some enterprise customers. The same customers are reporting that the model itself has gotten worse.
David Hsu, the founder of Retool, knows the feeling. Retool had been running Anthropic's Claude in production. The model was technically superior to the alternatives. But it kept going down. So Hsu switched to OpenAI. The inferior option, as he told the Wall Street Journal, because the inferior option stayed up.
Anthropic confirmed annualized recurring revenue of $30 billion in April, up from $9 billion at the end of 2025. The company appears to be using its leverage while it has it.
The performance degradation is not anecdotal. Stella Laurenzo, Senior Director of AI at AMD, analyzed 6,852 Claude Code session files and documented what she called a regression to the point where the model could not be trusted for complex engineering work. BridgeBench, an independent evaluation group, retested Claude Opus 4.6 and found it scored 68.3% accuracy, down from 83.3% in an earlier run and falling from second place to tenth. Paul Calcraft, a researcher who examined the comparison, noted the earlier result was based on six tasks and the retest on thirty, making direct comparison imperfect. The direction is not disputed.
On the infrastructure side, Anthropic's API uptime was 98.95% over the 90 days ending April 8, below the 99.99% industry standard. The company confirmed that cache time-to-live, the period during which repeated queries cost nothing, shifted from one hour to five minutes in early March. Boris Cherny, who leads Claude Code, said the company also moved the default effort setting to medium on March 3, 2026, citing user feedback about token usage.
Token usage across the industry is spiking regardless. OpenAI's token consumption rose from 6 billion per minute in October to 15 billion per minute by late March. Blackwell GPU rental prices climbed 48% in two months. CoreWeave raised prices more than 20% late last year. Anthropic is not alone in feeling the compute squeeze.
What is new is the combination: a price increase, a service decrease, and a company whose ARR grew 233% in a year telling enterprise customers to pay more for less.
Some customers will absorb the increase. The companies spending $1 million or more per year with Anthropic, of which there are now more than 1,000 according to the company, have fewer alternatives at the frontier. The switching cost is real. That is the bet.
Denise Dresser, OpenAI's revenue chief, alleged in a recent memo that Anthropic's run rate is inflated by around $8 billion via grossing up revenue share arrangements with Amazon and Google. Anthropic declined to comment on that allegation. Whether the ARR figure reflects cash or accounting treatment matters less to customers facing a renewal conversation this quarter.
The question for enterprise buyers is straightforward: premium pricing requires premium performance. The performance gap has narrowed. The price gap has widened. That arithmetic is prompting conversations that were not happening six months ago.
What to watch: whether Anthropic's large enterprise accounts, the ones generating the bulk of that $30 billion ARR, renegotiate or walk. The next renewal cycle will answer that question before any earnings call does.
