Claude Opus 5 Released — 1M Context, Adaptive Thinking Default, Complex Agentic Coding
Anthropic released Claude Opus 5 on July 24, 2026, for complex agentic coding and enterprise work. 1M-token context window, 128K max output, adaptive thinking enabled by default. Most-aligned model Anthropic has ever shipped per pre-deployment behavioral audit — lower deceptive behavior rates than Opus 4.8, Sonnet 5, or Fable 5.
Opus 5 is designed specifically for the longest-running agentic tasks — multi-hour autonomous sessions, large codebase refactors, complex enterprise workflows. Adaptive thinking default means the model decides its own reasoning depth per query. Notably: Anthropic explicitly avoided training Opus 5 on offensive cyber capability — the safety-forward positioning that also cleared it through faster government review than Fable 5 required.
API model string: claude-opus-5. Same schema as Opus 4.8. Recommended for: agentic sessions >2 hours, codebases >100K LOC, workflows requiring persistent context across many tool calls. For everything else, Sonnet 5 (introductory $2/$10 pricing through Aug 31) still wins on cost/performance.
Alignment note from the Opus 5 system card: automated behavioral audit found it “the least susceptible to being tricked into misuse” of any Claude model to date. In biology-research and offensive-security evaluations Anthropic conducted with private-sector and government partners, Opus 5 remains behind Mythos 5 in both categories — the intentional trade-off between capability and safety that’s now Anthropic’s public standard.