If you're on Anthropic's enterprise plan, you might have noticed the pricing page changed this week.
The old deal was straightforward: $200 per user per month, bundled with a block of API tokens. Predictable. Easy to budget for.
The new deal: $20 per seat plus full usage-based billing on every API call. A compliance expert quoted by The Register warned this "could potentially triple costs for some enterprise customers."
Why this matters more than you think
The bundled model was great if your team's usage was consistent. You paid $200, you got your tokens, you planned around it. The usage-based model punishes two things:
Unpredictable workloads. If your team has spiky usage (big refactors, new feature sprints, batch processing), your monthly bill becomes a surprise.
Model selection laziness. When tokens were bundled, nobody cared if a simple rename operation went through Opus. Now every call has a price tag and the expensive ones hurt.
What I'm doing about it
Three things have actually moved the needle for me:
Track before you optimize. You can't cut costs you can't see. I log every API call with the model, token count, and cost. Even a simple SQLite file with a rollup script gives you the visibility to know where money is going. Most teams discover that 60-70% of their calls are simple tasks hitting the most expensive model.
Match the model to the task. This is the big one. A docstring edit doesn't need Opus. A variable rename doesn't need Sonnet. The gap between Haiku's cost ($0.25/M input) and Opus ($15/M input) is 60x. If you can shift even 50% of your simple calls to a cheaper model, the math gets very different.
Set hard budget caps. Anthropic's API console lets you set monthly spending limits. Use them. Better to hit a cap and adjust than to get a surprise bill at the end of the month.
The bigger picture
This pricing shift isn't just Anthropic. OpenAI moved to usage-based pricing years ago. Google's Vertex AI is usage-based. The industry is converging on "pay for what you use" and that means cost management is now a core engineering skill, not an afterthought.
The teams that figure out intelligent model selection early will have a meaningful cost advantage. The ones that keep sending everything through the most expensive model will feel it in their budget.
Top comments (0)