Sign in
← AI PulseJul 26, 2026

Wire · news · Multi-source brief

Vercel AI SDK for Anthropic Updates Thinking Token Reporting

The Vercel AI SDK's Anthropic provider, @ai-sdk/anthropic@4.0.21, now reports 'thinking tokens' as 'reasoning token usage' as of July 26.

By Illumora Editorial

Source · Jul 26, 2026, 2:52 AM · On Illumora · Jul 26, 2026, 2:58 AM

Media from the primary source — shown here so you can stay on Illumora.

Synthesized from multiple allowlisted primaries on the same event. Lanes →

Brief drafted by Illumora’s editorial model from the linked primary source. Ops desk reviews flagged pieces. How we write →

Read the source →Vercel AI SDK Changelog — Release @ai-sdk/anthropic@4.0.21 · vercel/ai
Save

The Vercel AI SDK's Anthropic provider, @ai-sdk/anthropic@4.0.21, released on July 26, includes a patch that modifies how certain token types are reported. This update specifically addresses the reporting of 'thinking tokens' within the Anthropic integration.

Key Points

  • The @ai-sdk/anthropic@4.0.21 release occurred on July 26.
  • This update fixes how 'thinking tokens' are reported.
  • 'Thinking tokens' are now reported as 'reasoning token usage'.
  • The change is part of a patch release, e29788d.
  • Previous versions, including @ai-sdk/anthropic@4.0.20 and @ai-sdk/anthropic@3.0.102, introduced support for the Claude Opus 5 model.
  • Claude Opus 5 offers capabilities such as 128k output tokens, structured output, and adaptive thinking.
  • Earlier versions also added support for fallbacks in 'default' mode, routing safety classifier refusals to Anthropic's recommended fallback model.
  • Mid-conversation tool changes are supported via the toolChanges system message provider option, emitting tool_addition/tool_removal content blocks.

Context

According to the Vercel AI SDK changelog, the @ai-sdk/anthropic@4.0.21 update specifically addresses a reporting mechanism. This follows earlier updates to the Anthropic provider, including versions @ai-sdk/anthropic@4.0.20, @ai-sdk/anthropic@3.0.102, and @ai-sdk/anthropic@2.0.91, all released on July 24. These prior releases introduced significant features such as support for the Claude Opus 5 model, which includes advanced capabilities like 128k output tokens and adaptive thinking. They also added functionality for handling safety classifier refusals through a 'default' fallback mode and enabled mid-conversation tool changes.

Why It Matters

This update refines the accuracy of token usage reporting for developers using the Vercel AI SDK with Anthropic models. Accurate reporting of 'reasoning token usage' can assist in cost analysis and performance optimization for applications leveraging models that employ 'thinking tokens'.

What To Do

  • Developers using the Vercel AI SDK should note the change in token reporting for Anthropic models.
  • Review the updated documentation for @ai-sdk/anthropic to understand the implications of 'reasoning token usage'.
  • Compare token usage reports before and after updating to @ai-sdk/anthropic@4.0.21 to assess any impact on billing or analytics.
  • Test applications with the Claude Opus 5 model to evaluate its 128k output tokens and other frontier-tier capabilities.