Your AI now checks its own work. That changes your workflow, not your tool.

Anthropic shipped Opus 5 this week. The headline feature is not speed or price. It is that the model iterates until it succeeds instead of handing you a confident wrong answer.

Anthropic released Opus 5 on 24 July 2026, two months after Opus 4.8 landed on 28 May. The company's own framing is the interesting part. Opus 5 is described as "much stronger at verifying its work and iterating carefully until it succeeds."
In the demo that matters to builders, it wrote computer vision pipelines from prompts that were deliberately incomplete.
Three other things shipped in the same window:
• Anthropic's safety classifiers are expected to trigger roughly 85% less often on Opus 5 than on the previous generation, and a new Automatic Fallbacks beta routes a blocked request to a smaller model instead of failing outright. Fewer dead ends mid task.
• Claude voice mode stopped being Haiku only. It now runs Opus, Sonnet or Haiku, defaults to the fastest version of whatever you last used in text, speaks ten languages, and connects to Gmail, Calendar, Slack, Canva and Notion. Free accounts stay on Haiku with one connected app.
• On the other side of the fence, OpenAI's frontier models and Codex went generally available on Amazon Bedrock, and Microsoft's MAI-Code-1-Flash reached GitHub Copilot Business and Enterprise.
Stop reading that as a shopping list
The instinct when four things ship in a week is to go switch tools. Do not. Switching costs you a day and buys you almost nothing, because within a quarter every serious model will have the same features.
The thing worth changing is how you brief the model.
The actual workflow change
If a model is now genuinely better at verifying its own work, the highest leverage thing you can write in a prompt is the test, not the instruction.
Old habit: describe the task in detail and hope. New habit: describe the task briefly, then define exactly what "done and correct" looks like, and tell it to keep going until it passes.
That is it. That is the whole upgrade. You are moving from giving directions to setting an acceptance criteria, and the model closes the gap itself.
One caveat, stated plainly
Anthropic did not publish specific benchmark scores in the launch coverage, and pricing was not disclosed in it either. "Stronger at verifying its work" is the vendor's claim. Test it on your own workload before you rebuild anything around it. That is not cynicism, that is how you avoid rewriting your stack twice.