Upgrade Claude Code From Opus 4.6 → 4.7 in 5 Minutes

OK so here’s what’s actually going on with Opus 4.7

Anthropic quietly dropped claude-opus-4-7, and if you’re living in Claude Code all day, this is basically a brain transplant for your dev environment.

It’s smarter, pickier, more expensive on tokens, and slightly more paranoid about cyber stuff.

Let’s walk through the upgrade like you’re migrating apartments, not rewriting your life.

1. Swap the model ID: claude-opus-4-6 → claude-opus-4-7

This is the “change your address with the post office” step.

1. Find everywhere you hard‑coded the model. Anywhere you see claude-opus-4-6 in scripts, CI configs, editor plugins, or Claude Code settings, that’s your target.
2. Swap it to the new ID. The new hotness is:

   claude-opus-4-7

   No “preview”, no extra suffix, nothing cute.
3. Smoke test a simple command. Ask it something obvious like “Which model are you?” If you don’t see 4.7 mentioned in the self‑report or system info, you missed a config somewhere.

2. Update your settings.json (Claude Code)

This is the one that bites people, because you set it once and then totally forget it exists.

In Claude Code, crack open your config and find the model entry.

Here’s the before → after you probably need:

{
  "claude": {
    "model": "claude-opus-4-7",
    "effort": "high",
    "maxTokens": 4000
  }
}

1. Make sure the model field is exactly claude-opus-4-7. Typos here turn into super confusing “why is this so dumb today?” vibes.
2. Keep an eye on maxTokens. 4.7 “thinks harder” at higher effort levels, which usually means more output tokens. If you were already flirting with your limit, bump it a bit or be ready for truncation.

Restart your editor / Claude Code session after changing this, or it may keep the old model in memory.

3. CLI flag: use claude --model claude-opus-4-7

Command line people, this is your line in the sand.

claude --model claude-opus-4-7 \
       --effort high \
       --file src/main.py

1. Search your scripts for old flags. Anywhere you scripted claude --model claude-opus-4-6, swap it to 4-7.
2. Standardize on one spot. If you use a wrapper script (like dev-cli.sh), define the model there once so you’re not fixing this in 8 places six months from now.

4. The new xhigh effort level: when to use it

Effort in Claude Code is basically “how hard should this thing sweat over your problem?”

We now have a new gear: xhigh, sitting between high and max.

1. Use high for everyday coding. Refactors, writing new modules, test generation, quick bug hunts. Good balance of speed, cost, and quality.
2. Use xhigh for “this can’t be wrong.” Examples:
   • Tricky concurrency / async logic.
   • Cross‑service data flows (auth, billing, PII).
   • Schema migrations or dangerous data transforms.
   • Non‑trivial architectural decisions (“Should we split this into a separate service?”).
   Think of it as “staff engineer review level” without going full send.
3. Reserve max for true boss fights. Huge legacy rewrites, mixed‑language codebases, complex financial logic, or “I’d pay a consultant for this” questions. max will think even harder, which = more time and tokens.

In settings.json, it’s literally:

"effort": "xhigh"

5. The new /ultrareview command: your fake principal engineer

/ultrareview is like calling in the grumpy senior dev who actually reads the whole PR.

When you run /ultrareview in Claude Code on a file or diff, 4.7 goes into “senior reviewer” mode:

• Flags subtle design flaws, not just syntax or style.
• Calls out logic gaps, edge cases, race conditions, weird failure modes.
• Comments on architecture-level issues, not just line‑by‑line nits.

1. Use it before merging critical changes. Stuff like auth flows, payment code, data deletion, infra scripts, anything where a silent bug = real money or real pain.
2. Use it on “mysterious” flaky areas. That one module nobody wants to touch? Run /ultrareview on it and see what it hates.
3. Don’t spam it on every trivial PR. It’s slower and more expensive than a quick inline answer. Treat it like a serious code review, not a linter.

6. GOTCHA: tokenizer change = higher token counts, same prompts

Here’s the painful part.

Opus 4.7 has an updated tokenizer. That means the same exact prompt will usually count as more input tokens than on 4.6. Roughly:

• 1.0x – 1.35x the input tokens vs 4.6, depending on your text.

API pricing stayed at $5 / $25 per million (input / output), but if you’re eating 1.2x tokens per call, your bill follows.

1. Do a small A/B test. Run a representative workload on 4.6, note total tokens. Run the exact same prompts on 4.7, compare. Multiply that delta by your monthly volume — that’s your new bill.
2. Watch output tokens too. 4.7 “thinks harder” at high/xhigh/max, which typically means longer answers. If you’re not careful, you pay twice: more input, more output.

7. When NOT to upgrade (yet)

This is the part nobody likes to say out loud, but here we are.

You probably should not upgrade immediately if:

• You’re super token‑budget‑constrained.
• You run a high‑volume workflow (thousands/millions of calls per month).
• Your current 4.6 setup is “good enough” and mostly doing autocomplete / simple refactors.

1. Run the math first. If 4.7 bumps you from, say, $3k → $3.6k/month and that’s not in the budget, stay on 4.6 where it’s safe.
2. Be selective about 4.7 usage. Common pattern:
   • Keep 4.6 for cheap, bulk flows (small edits, rote transformations).
   • Use 4.7 only for complex reviews, architecture questions, or /ultrareview.
3. Dial effort down if you do upgrade. If you move to 4.7 and your wallet screams, try medium or high instead of xhigh/max for routine stuff.

8. Cybersec safeguards: why some pentest prompts now get blocked

Opus 4.7 is the first Anthropic model with automated systems to detect and block some cybersecurity prompts.

If you’re doing legit offensive security, you will run into this.

• Exploit crafting, payload generation, vuln chaining: more likely to get blocked.
• General secure‑coding advice: usually fine.

1. Don’t assume 4.7 is “broken” when it refuses. A blocked response here is the new normal, not a bug.
2. If you’re a real pentester / vuln researcher, get verified. Apply for Anthropic’s Cyber Verification Program. That’s the official path if your job is literally to break things and you need the model to cooperate.
3. Keep 4.6 around if you’re mid‑engagement. If you’re in the middle of a client pentest and don’t have verification yet, don’t flip everything to 4.7 overnight.

Bottom line: should you move your Claude Code life to 4.7?

Opus 4.7 is Anthropic’s most capable generally available model right now, and they’re openly saying it’s beating GPT‑5.4 and Gemini 3.1 Pro on the stuff devs care about: agentic coding, scaled tool use, computer control, financial logic.

It also does better vision (up to 2,576px on the long edge), which matters if you’re tossing it screenshots of UIs, dashboards, or code snippets in images.

Verdict: if quality and safety matter more than raw token cost, upgrade to claude-opus-4-7, lean on xhigh + /ultrareview for the scary work, and keep a calculator handy for your token budget. If you’re running a razor‑thin, high‑volume pipeline, do the math first and consider a hybrid 4.6 + 4.7 setup.

Now you know more than 99% of people.

Now you know more than 99% of people. — Sara Plaintext