What's New in Claude Opus 4.7

Anthropic just shipped Opus 4.7. It's the biggest jump Claude Code has ever made. I've been running it on real projects for a few days and the difference is night and day. Faster reasoning. Sharper code. A 1 million token context window that changes how you actually build. Not even exaggerating, this is the upgrade everyone's been waiting for.

Here is what it means for you. Bigger projects without Claude losing the plot. Long agentic tasks that run for hours and actually finish. Output quality that jumped across the board compared to Opus 4.6, which was already absurd. If you only care about one Claude Code upgrade this year, this is the one.

But here is the thing most people miss. Switching the model is the easy part. Every creator on the internet will tell you that. The real unlock is in how you use it. Boris, the guy who runs Claude Code at Anthropic, just posted six tips after dogfooding 4.7 for weeks. These genuinely 2-3x what I get out of Claude. Every one is in this guide.

What model are you running right now? Give me the exact model ID and your context window size. Then tell me what you can do in this conversation that you could not do on Opus 4.6.

Now the real stuff. These six things are what make 4.7 feel like a different tool. Not an upgrade. A different tool.

1. Auto mode. Opus 4.7 loves long-running tasks. Deep research. Big refactors. Building entire features. Iterating until it hits a performance benchmark. In the past, you either had to babysit every permission prompt or use --dangerously-skip-permissions, which is scary. Auto mode is the safer middle. Every permission request gets sent to a classifier that decides if the command is safe. If it is, it auto-approves. No more babysitting. You can run multiple Claudes in parallel. Once one is cooking, switch to the next.

2. The new /fewer-permission-prompts skill. If you don't want auto mode, this one is gold. Run it and Claude scans your session history for safe bash and MCP commands you keep approving. Then it hands you a list of permissions to add to your allowlist permanently. Takes 30 seconds. Saves you hundreds of clicks a week.

/fewer-permission-prompts

3. Recaps. Opus 4.7 ships with recaps built in. Every long-running session now gets a short summary of what Claude just did and what's next. You walk away for lunch, come back two hours later, and you're not lost. The recap tells you exactly where things stand. Tiny feature, massive quality-of-life upgrade.

4. Focus mode. This is the one I keep coming back to. Focus mode hides all the intermediate tool calls and shows you the final result. Claude runs commands, edits files, verifies its own work, and you just see the outcome. Toggle it with /focus. Use it when you trust Claude to handle the details, which with 4.7 is most of the time.

/focus

5. Configure your effort level. Opus 4.7 uses adaptive thinking instead of fixed thinking budgets. You don't set a token limit anymore, you set an effort level. Lower effort means faster and cheaper. Higher effort means more intelligent. Boris uses xhigh for most tasks and max for the hardest ones. I've been copying him and the difference on complex work is obvious.

/effort

6. Give Claude a way to verify its work. This is the most important one on the entire list. When Claude can test itself end to end, the output quality jumps another 2-3x. Boris lives this so hard he built a custom skill called /go. His prompts literally look like "Claude do blah blah /go" and the skill makes Claude test end to end, run /simplify, then open a PR. I'm stealing this pattern.

Add [feature] to this project. When you are done, do these three things in order:

1. Test it end to end. Start the dev server if it's a frontend change, hit the endpoint directly if it's backend. Confirm it actually works before you stop.
2. Run /simplify on the changes to strip out anything overcomplicated or unnecessary.
3. Open a pull request with a clear title and description.

Do not stop until all three are done.

And that 1 million token context window is still sitting underneath all of it. Five full novels worth of text in one conversation. Your entire codebase, every doc, every README, your product spec, all in Claude's head at once. Before this you had to trim, summarize, or split work across chats. Now you just paste everything in and keep going.

That's the whole thing. Switch to Opus 4.7. Turn on auto mode. Run /fewer-permission-prompts. Use /focus when you trust it. Crank your effort with /effort. And always give Claude a way to verify. Do those six things and 4.7 stops feeling like an upgrade. It starts feeling like a different tool entirely.

The full course walks through my exact Claude Code workflow with Opus 4.7. How I plan. How I prompt. How I verify. And the exact file that makes Claude build production-quality apps from the very first prompt.