AINEWS
Anthropic prices Claude Sonnet 5 for the AI agents burning enterprise budgets
Agentic AI, the multi step autonomous systems that fire hundreds of API calls per task, is the workload driving enterprise token bills.
Agentic AI, the multi step autonomous systems that fire hundreds of API calls per task, is the workload driving enterprise token bills.
Anthropic has a new model for the workload that makes enterprise AI expensive: the multi-step autonomous agents that fire hundreds or thousands of API calls per task and run up token bills most enterprises cannot audit, let alone contain. Claude Sonnet 5, released Tuesday, is priced for that workload rather than for chatbot-style single-prompt queries.
Sonnet 5 is Anthropic's mid-tier Claude line. The new model's standard pricing — taking effect September 1, 2026 — is $3 per million input tokens and $15 per million output tokens, compared with $4 and $25 for the higher-end Claude Opus 4.8. Through August 31, 2026, Anthropic is running an introductory price of $2 per million input tokens and $10 per million output tokens, half the input cost of Opus 4.8 and 60% less on output. The agentic-workload framing matters because "agentic AI" refers to systems, often coding or research agents, that take multi-step actions across tools and APIs rather than answering a single question, so query volume and token cost compound across long task chains in a way that simple chat queries do not.
Anthropic's positioning, detailed on the Sonnet 5 launch page and the Sonnet product page, is that the model delivers performance closer to Opus-tier on the high-query, agentic tasks where enterprise customers most often blow token budgets. The company is not claiming Opus parity across every benchmark, only that the cost-versus-performance tradeoff has shifted for the specific workload class where token math has historically broken down. The model's documentation entry and system card describe the capabilities in more detail.
The distribution choices reinforce the framing. Sonnet 5 is the new default model across Anthropic's free, Pro, Claude Code, and Claude Platform surfaces, which means builders evaluating a high-volume agentic loop can experiment on the free tier before committing to a paid plan.
The caveat Anthropic itself surfaces is real. Sonnet 5's lower per-token price may not prevent over-spending in every case: agents that loop, retry, or fan out badly can still burn budget regardless of the model rate, and the cost-control problem shifts from per-call price to per-task architecture. What changes is the floor. The workload that used to require either an Opus-tier budget or a hand-rolled cheaper-model workaround now has a model that is explicitly priced and tuned for it.
What to watch: whether the introductory pricing window, which closes on August 31, holds in practice as enterprise contracts are signed, and whether benchmark-by-benchmark comparisons of Sonnet 5 against Opus 4.8 on agentic tasks confirm the "closer to Opus" framing Anthropic is using.