SonarSource's independent evaluation explicitly names "Claude Opus 5 Thinking (adaptive thinking mode)" and benchmarks it against Opus 4.8 Thinking on a Java task suite of 4,441 problems spanning HumanEval, MBPP, and ComplexCodeEval. The TLDR figures: an 88.6% pass rate across the 544 HumanEval and MBPP executable-test The review also documents a volume shift: Opus 5 generates 2.3 times as much code as Opus 4.8, with total findings rising 2.7 times — Sonar frames this as changing what "verification" looks like in practice, since the self-verification Anthropic highlights pays off in higher correctness despite larger output. The piece

Anthropic's first-party documentation introduces Claude Opus 5 as a step-change over Opus 4.8, with the largest gains in deep reasoning, agentic/long-horizon work, and test-time compute scaling. The model ships with a 1M-token context window (default and maximum, with no smaller variant), 128k max output tokens, thinki Several platform features ship alongside the model: a mid-conversation tool changes beta (header mid-conversation-tool-changes-2026-07-01) that lets developers add or remove tools between turns while preserving the prompt cache; a new "default" mode for the fallbacks parameter (beta, header server-side-fallback-2026-07