What changed
On June 26, Anthropic said it is increasing rate limits across all Claude model tiers and simplifying how customers move through capacity bands. The company is replacing its older tier structure with three clearer levels: Start, Build, and Scale. Anthropic also said the update includes higher limits for Claude Sonnet 4.5 and a change to how prompt caching tokens are counted inside rate limits. Existing customers keep their current limits until September 1, 2026 unless they ask sales or support to move earlier. Source: Anthropic API release notes for June 26, 2026.
Why it matters
This is a product update, not just a policy tweak. One of the most common failure points in AI products is that early prototypes work, then real usage runs into request and token ceilings that are hard to predict or negotiate around. By raising limits and collapsing the upgrade path into three named tiers, Anthropic is making Claude capacity easier to reason about for teams that are moving from experiments to production. The official Claude API rate limits documentation makes clear that Anthropic now wants builders to think in terms of request, input-token, and output-token budgets instead of opaque account states.
What builders should do now
If you ship on Claude, this is a good week to revisit your throughput assumptions. First, check whether your workloads would benefit immediately from the higher Sonnet 4.5 limits or the simpler Start, Build, and Scale progression. Second, review any prompt-caching-heavy flows, because Anthropic now explicitly counts prompt caching tokens inside rate limits, which can change how much headroom cached workflows really have. Third, if you are an existing customer on the legacy setup, decide whether to stay put until the automatic September 1 migration or request earlier access.
The broader takeaway is that model competition is now happening in operational UX as much as raw model quality. Better reasoning models matter, but clearer scaling rules are what let product teams ship them confidently.