Setting Up Opus 4.7 in Cursor, Zed, the API, Bedrock, and Vertex

Claude Opus 4.7 is basically the “I hired a senior staff engineer and a hedge fund quant in one” model.

Anthropic says it’s beating GPT‑5.4 and Gemini 3.1 Pro on coding, tool use, computer control, and finance.

So let’s wire this thing into all your tools without breaking anything.

1. Claude Code CLI (claude-opus-4-7)

If you’re already using Claude Code, this is just a model swap.

Quick CLI flag setup:

claude --model claude-opus-4-7

Or if you’re using a config file (usually ~/.claude/settings.json or similar):

{
  "default_model": "claude-opus-4-7",
  "effort": "xhigh"
}

That "effort": "xhigh" is the new sweet spot: smarter than high, not as slow/wordy as max.

New power move: /ultrareview

/ultrareview
# Paste your PR diff or file here

This runs the new senior-reviewer mode, which is way better at catching “this design bites us in 6 months” problems than “you missed a semicolon.”

What to expect: noticeably deeper code reviews and more “are you sure this is the right abstraction?” style feedback, plus longer outputs if you crank effort too high.

2. Cursor (custom model: claude-opus-4-7)

Cursor makes this pretty painless.

In the app:

Open Settings → Models
Go to Anthropic or Custom models section
Add a custom model with ID claude-opus-4-7

Rough JSON-style config if Cursor asks for it:

{
  "name": "Claude Opus 4.7",
  "provider": "anthropic",
  "model": "claude-opus-4-7"
}

Then set it as default for:

Chat
Inline code completions (if allowed)
Test generation / refactors

What to expect: better multi-file refactors, more reliable tool-calling, and Cursor agents that feel a lot more like a real pair‑programmer than a code autocomplete toy.

3. Zed (assistant.json → claude-opus-4-7)

Zed uses a little JSON file to define your AI “assistant.”

Find or create .zed/assistant.json in your project or user config.

{
  "assistant": {
    "provider": "anthropic",
    "model": "claude-opus-4-7",
    "effort": "xhigh",
    "temperature": 0.2
  }
}

Some Zed setups use a slightly different schema, but the key is the "model": "claude-opus-4-7" line. If Zed complains, check their docs for any new required fields.

What to expect: smarter “read the whole repo and help me” behavior; it’s much better at multi‑step edits and not losing the plot halfway through a refactor.

4. Anthropic API directly (raw requests)

This is where you talk straight to @AnthropicAI, no middleman.

Messages API (recommended):

POST https://api.anthropic.com/v1/messages
Content-Type: application/json
x-api-key: YOUR_API_KEY
anthropic-version: "2023-06-01"

{
  "model": "claude-opus-4-7",
  "max_tokens": 2048,
  "effort": "xhigh",
  "messages": [
    {
      "role": "user",
      "content": "Explain this code like I'm new to the repo..."
    }
  ]
}

That "effort": "xhigh" flag is new for Opus 4.7; it makes the model “think harder” but not go full essay mode like max often does.

What to expect: same $5 / $25 per million token pricing as 4.6, but with more tokens consumed because the tokenizer changed and the model tends to output more when effort is high.

5. Amazon Bedrock (anthropic.claude-opus-4-7-v1:0)

This one’s slightly messy because AWS loves region-specific weirdness.

Anthropic’s naming pattern for 4.7 on Bedrock is:

anthropic.claude-opus-4-7-v1:0

Important: double-check this in the Bedrock console for your region (some regions sometimes use -v1 vs -v1:0 variants).

Example InvokeModel payload:

{
  "modelId": "anthropic.claude-opus-4-7-v1:0",
  "contentType": "application/json",
  "accept": "application/json",
  "body": {
    "model": "claude-opus-4-7",
    "max_tokens": 2048,
    "effort": "xhigh",
    "messages": [
      {
        "role": "user",
        "content": "Help me analyze this trading strategy..."
      }
    ]
  }
}

Some SDKs wrap that differently (e.g. inputText vs raw anthropic-style body), so treat this as the “shape,” not a copy‑paste for every language.

What to expect: Opus‑level reasoning running inside your AWS stack, but be prepared for slightly higher token counts vs 4.6 because of the new tokenizer.

6. Google Vertex AI (publisher: anthropic, model: claude-opus-4-7)

On Vertex, you pick the model by publisher + name in the UI, or by a resource name in code.

Console path looks roughly like:

Go to Vertex AI → Model Garden
Filter by Publisher: anthropic
Pick claude-opus-4-7

Typical full model name pattern (this part does vary by region/project):

projects/YOUR_PROJECT_ID/locations/us-central1/publishers/anthropic/models/claude-opus-4-7

So a JSON-ish request via the Vertex SDK ends up like:

{
  "model": "projects/YOUR_PROJECT_ID/locations/us-central1/publishers/anthropic/models/claude-opus-4-7",
  "contents": [
    {
      "role": "user",
      "parts": [{ "text": "Design a robust API rate limiter..." }]
    }
  ],
  "params": {
    "max_tokens": 2048,
    "effort": "xhigh"
  }
}

Honestly, even I’d double-check the exact resource string in the Vertex console because Google loves subtle naming differences per region.

What to expect: Anthropic brains, Google infra; good if your security team only blesses GCP but you still want the Opus 4.7 reasoning and vision upgrades.

7. Microsoft Foundry (catalog: anthropic-claude-opus-4-7)

Foundry (inside Azure) exposes Anthropic models through its own catalog IDs.

For Opus 4.7, that’s:

anthropic-claude-opus-4-7

In a Foundry deployment JSON, it’ll look something like:

{
  "deploymentName": "opus-4-7-prod",
  "model": {
    "catalogId": "anthropic-claude-opus-4-7"
  },
  "inference": {
    "parameters": {
      "max_tokens": 2048,
      "effort": "xhigh"
    }
  }
}

The exact wrapper fields vary a bit across Azure OpenAI / Foundry flavors, so treat "catalogId": "anthropic-claude-opus-4-7" as the key line and confirm the rest in your portal.

What to expect: Opus 4.7 plugged into your Microsoft stack, with the same strengths: agentic coding, tool use, financial analysis, and better vision.

Common gotchas across all tools (read this before your CFO yells)

Tokenizer changed → token counts jump.
Opus 4.7 uses a new tokenizer that often counts 1.0–1.35x more input tokens than 4.6 for the same text. Translation: same price per million tokens, potentially bigger bill if you just swap models and keep stuffing in huge contexts.
xhigh effort = more thinking, more tokens.
The new xhigh level sits between high and max. It’s great for deep reasoning, but the model tends to “think out loud” more, which means longer outputs and higher output token usage. Watch your max_tokens limits.
Stricter cybersecurity safeguards.
Opus 4.7 is the first Claude model with automated systems to catch and block prohibited cybersec content. So “help me secure this system” is fine, “write me an exploit for X” will (correctly) get blocked unless you’re in Anthropic’s new Cyber Verification Program as a legit pentester / vuln researcher.

Bottom line: plug in claude-opus-4-7 everywhere, set effort to xhigh where you care about quality, trim your prompts a bit to offset the new tokenizer, and you’re running one of the strongest public models on the market.

Now you know more than 99% of people.

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