What Changed in Sonnet 4
Claude Sonnet 4, released July 15, 2025, represented the largest single-generation improvement in the Sonnet line since Claude 3.5 Sonnet established itself as the developer favourite in 2024. The headline improvements were in software engineering — SWE-bench Verified score improved from 62.3% (Sonnet 3.7) to 72.7%, meaning Sonnet 4 correctly resolved nearly three-quarters of real-world GitHub issues in automated testing.
Instruction following improved substantially on Anthropic's internal IFEval benchmark, with the model better at adhering to complex multi-part instructions, format constraints, and output length specifications. This was practically significant for developers building agents that required structured JSON output, specific response templates, or constrained generation formats.
Agentic Task Completion
Sonnet 4 showed the most dramatic improvements in agentic contexts — multi-step tasks where the model must plan, execute tool calls, evaluate results, and iterate. On TAU-bench (testing agentic performance across realistic business tasks), Sonnet 4 scored 88.4% versus Sonnet 3.7's 81.2%. This 7-point improvement was significant, as TAU-bench tests the kinds of real-world business automation tasks that most enterprise customers actually care about.
Pricing and Availability
Anthropic maintained Sonnet 4 at the same pricing as Sonnet 3.7: $3.00/million input tokens, $15.00/million output. Sonnet 4 became the default model on Claude.ai for all users and the default recommendation in the Anthropic API documentation. Both Opus 4 (for highest capability tasks) and Haiku 3.5 (for speed and cost-sensitive tasks) remained available alongside Sonnet 4 in the API.
Claude Code Integration
Sonnet 4 immediately became the default model powering Claude Code, Anthropic's terminal-based coding agent. Users of Claude Code reported noticeable quality improvements: fewer rounds of correction needed, more accurate implementation of complex features, and better code that followed existing codebase conventions without explicit instruction to do so.
What This Means for Indian Businesses
Claude Sonnet 4 is the model powering Claude Code — Anthropic's terminal-based coding agent that Indian developers are increasingly using to accelerate software delivery. Its improvements in instruction following mean fewer correction loops and more reliable output on the first attempt. For Indian software agencies billing by the project rather than the hour, Claude Sonnet 4's increased reliability directly improves profitability.