Claude Web Search Filtering: Cheaper Cited Agents in Python

Anthropic shipped Claude Sonnet 5 on June 30, 2026, and by July 1 it was the default model for every Free and Pro user. The headline was not a benchmark. It was the price: $2 per million input tokens and $10 per million output through August 31, 2026, for a model that plans, browses, and runs tools close to the level of Opus 4.8. Anthropic pitched it as a cheaper way to run agents, and that framing landed at exactly the moment teams were staring at Q2 agent bills they did not expect.

Here is the part that does not fix itself when the model gets cheaper. If you build a research agent that searches the web, the raw pages it pulls back are dumped into the context window as input tokens on every turn that follows. A single question can trigger ten searches, and each result page can be thousands of tokens. You pay for all of it, again and again, as the conversation grows. A cheaper model with an expensive context is still an expensive agent.