# Anthropic releases Claude Sonnet 5 for agentic workloads > Sonnet 5 adds effort controls and a revised tokenizer across Claude plans and the API, making migration tests more important than alias swaps. Canonical URL: https://www.devobs.io/news/news-claude-sonnet-5-release/ By: Claire Dubois Published: 2026-09-06T11:58:54.625Z Updated: 2026-09-06T11:58:54.625Z Event date: 2026-06-30 Section: AI Anthropic released Claude Sonnet 5 on June 30 across Claude plans, Claude Code, and the Claude API. The [launch announcement](https://www.anthropic.com/news/claude-sonnet-5) emphasizes planning, browser and terminal tool use, coding, and longer agentic execution. Developers can address the model with the `claude-sonnet-5` identifier and select effort levels for different workloads. ## The tokenizer makes cost migration nontrivial Anthropic says Sonnet 5 uses an updated tokenizer and that the same input can map to roughly 1.0 to 1.35 times as many tokens depending on content. Pricing in the post was also edited after launch, and an evaluation chart was corrected to use Anthropic's standard methodology. Those changes are reasons to store the source date and changelog with any procurement or benchmark decision. A model can become cheaper per token while making a particular prompt longer in tokens. Capacity planning should therefore use full request traces, including system instructions, retrieved context, tool results, compaction, retries, and output. ## Agentic claims require end-to-end evaluation Anthropic reports gains over Sonnet 4.6 and compares higher effort settings with Opus 4.8. These are vendor evaluations and early-user reports. For an application, the useful unit is a completed task under its own tool and review constraints. Run a shadow evaluation with pinned model IDs. Measure accepted outcomes, human corrections, tool-call errors, total tokens, elapsed time, and unsafe or out-of-scope actions. Repeat tasks to expose variance, and test low through high effort rather than assuming the maximum is always economical. Update budgets and rate-limit alarms only after observing the new tokenizer on production-shaped context. The migration should end with an explicit routing rule for which tasks belong on Sonnet 5 and which still justify a different model. ## Source references -