Descriptions:
Sam Witteveen offers a contrarian take on Anthropic’s Claude Sonnet 4.6 release, arguing that the widely celebrated price reduction obscures a serious token consumption problem that could make the model more expensive in practice than its predecessor. While Sonnet 4.6 is priced 40% below Opus 4.6 on a per-token basis and extends context to one million tokens, Witteveen points to independent benchmarks from Artificial Analysis showing the model used 280 million tokens on their evaluation suite — compared to 58 million for Sonnet 4.5 and 160 million for Opus 4.6.
The culprit appears to be adaptive thinking, the feature that allows Sonnet 4.6 to dynamically apply extended chain-of-thought reasoning. Witteveen notes this is the same pattern that caused early GPT-5 deployments to generate unexpectedly high API bills, and argues that practitioners should run their own token-per-task benchmarks before assuming the cheaper nominal price translates to lower real costs.
The video also raises a second concern: API feature parity is breaking down across cloud providers. Programmatic tool calling with server-side code execution — available natively on some platforms — is not uniformly supported across Anthropic direct, Google Cloud, and AWS deployments of the same model, meaning the effective capability of Sonnet 4.6 varies depending on where it is accessed. Developers building agentic workflows or cost-sensitive production systems will want to factor both issues into their model selection decisions.
📺 Source: Sam Witteveen · Published February 18, 2026
🏷️ Format: Review
