Anthropic reframes model pricing ladder

The meaningful metric is cost per completed task, not sticker price; developers may soon route workloads by capability, latency, and efficiency rather than simply choosing the cheapest or strongest tier.

• –Anthropic describes Claude models along a capability-cost-speed curve, not a rigid hierarchy.
• –Cheaper high-quality models make agent loops, batch coding, and background workloads more practical.
• –Production teams should benchmark total task cost, including retries, tool calls, tokens, and latency.
• –List prices and usage limits still differ sharply, so efficiency gains must be validated against real workloads and not just roadshow claims.