Disclosure: Some links below are affiliate links. If you buy through them, this site earns a small commission at no extra cost. Editorial recommendations are never influenced by affiliate rates.

Anthropic dropped Claude Opus 4.7 on 16 April 2026 with gains across coding and vision tasks. On SWE-bench Verified (the standard benchmark for agentic coding), it scores 87.6% compared to 80.8% on Opus 4.6. Vision accuracy jumped from 54.5% to 98.5%. The pricing stayed flat: $5 per million input tokens, $25 per million output tokens. (Source: anthropic.com/news/claude-opus-4-7)

Whether it's worth upgrading depends on what you're doing with it. Here's the practical version.

What changed in 4.7

Coding got a lot better

The coding gains are the headline. Opus 4.7 scores 70% on CursorBench (a real-world IDE-integrated benchmark), up from 58% on Opus 4.6. That 12-point gain measures multi-file refactors and context-aware completions, which is what developers actually do. (Source: nxcode.io/claude-opus-4-7-guide)

On SWE-bench Verified (the agentic coding standard), 4.7 scores 87.6% versus 80.8% for Opus 4.6. GPT-5.4 sits at ~70% on the same benchmark. Clear difference.

In practice, tasks that previously required close human supervision (debugging a codebase, writing a fix, running tests) now run more independently. One early tester (a fintech firm) noted: "It catches its own logical faults during the planning phase and accelerates execution." (Source: anthropic.com/news/claude-opus-4-7)

Vision is now actually useful

The vision jump is the most dramatic change. Image resolution went from 1.25MP to 3.75MP. Accuracy on visual-understanding tasks went from 54.5% to 98.5%. The coordinate system simplified to 1:1 mapping.

Practically: you can now feed Claude screenshots for QA, UI debugging, chart reading, and document review, and it will get them right. The old model struggled with dense tables or small fonts. This one doesn't. If you've been skipping vision tasks, this is worth trying again.

New control features

Three changes for API users:

xhigh effort level. New step between "high" and "max" for extended thinking. Use when you need more reasoning than "high" but want to avoid "max" cost ceiling.

task_budget (beta). Instead of a hard per-request token limit, you now set an approximate budget for the full agentic loop: tool calls, thinking, output. The model winds down gracefully at the limit instead of cutting off mid-sentence. Makes long document reviews simpler to cost-estimate.

/ultrareview command in Claude Code. Deeper code review in the Claude Code editor. Relevant mainly for the desktop app, not the API.

What it costs (and the tokeniser catch)

The per-token price matches Opus 4.6: $5 per million input tokens, $25 per million output tokens. (Source: cloudzero.com/claude-api-pricing)

The wrinkle: 4.7 includes a new tokeniser. Depending on your content, the same text uses 1.0x to 1.35x more tokens. If you run tight cost projections on batch workflows, you need to retest before switching. You might pay 0 to 35% more per request even at unchanged per-token rates. (Source: finout.io/claude-opus-4-7-pricing)

For NZ developers: 5/25 per million tokens translates to NZD $8.25 input / $41.25 output at 1:1.65 exchange rate. Prompt caching cuts input costs up to 90%. Batch processing cuts both by 50%. Most small-to-medium workloads cost less than a coffee per month.

Or skip the API entirely and use Claude Pro for $20 USD/month (~NZD $33) via claude.ai. You get full-model access with conversation limits, no API setup. Better if you're an individual or small team that doesn't need programmatic access.

Alternatively, Perplexity Pro (20USD/month)bundlesClaudeOpus4.7accessalongsidewebsearchandimageanalysis, whichsavesswitchingbetweentools.Or[ReplitFoundingTier](https://replit.com/pricing)(20 USD/month) includes Claude Opus for pair-programming and rapid prototyping.

How it compares to the competition

Model SWE-bench Verified SWE-bench Pro Computer Use
Claude Opus 4.7 87.6% 64.3% 78.0%
GPT-5.4 ~70% 57.7% 75.0%
Gemini 3.1 Pro lower lower lower

Opus 4.7 leads on every publicly available agentic coding and computer-use benchmark. (Source: llm-stats.com/claude-opus-4-7-launch, vellum.ai/claude-opus-4-7-benchmarks)

Note: Anthropic has Claude Mythos Preview, which scores higher on most benchmarks. It's not publicly available yet. Anthropic is rolling it out slowly to test cybersecurity safeguards. Opus 4.7 is what you can use today.

What Anthropic changed around cybersecurity

Opus 4.7 is the first Claude with real-time automated safeguards for cybersecurity requests. This is Anthropic's testing ground for eventual broader model releases. (Source: support.claude.com/en/articles/14604842)

Security pros doing legitimate work (penetration testing, vulnerability research) can apply to Anthropic's Cyber Verification Program for less-restricted access.

For most business users, this doesn't matter. You won't hit the safety layer building coding agents, document workflows, or analytics. But if you're in a compliance-heavy industry, the safeguards exist and are documented.

Breaking changes to watch if migrating from 4.6

Three API changes that might break existing code:

If your prompts relied on specific sampling behaviour, test before rolling out.

Where to access Opus 4.7

Access Opus 4.7 through:

No NZ restrictions. USD billing is standard everywhere.

TipClaude Pro ($20 USD/month, ~NZD $33) gives chat access to Opus 4.7. For API use with per-token billing, sign up at platform.claude.com. If you want Claude bundled with web search and other tools, Perplexity Pro is a good alternative. Billing is USD only.

Who should actually upgrade

Upgrade now if:

Test carefully if:

Stick with Sonnet 4.6 if:

For most NZ small businesses using Claude for writing, email, or support summaries, Sonnet 4.6 is still the right pick. Opus 4.7 pays off when tasks are hard and multi-step. If your prompts run 30 seconds and produce short text, you're paying 5x for capability you won't use.

For developers building agent workflows, though, Opus 4.7 is the best publicly available option right now.


Sources: Anthropic launch post (April 16, 2026), nxcode.io complete guide, evolink.ai review, llm-stats.com benchmark analysis, vellum.ai benchmark explainer, finout.io pricing analysis

TD
Toby Downs is an independent tech writer based in New Zealand, covering SaaS, AI tools, and business software for tpdowns.com. No paid placements, no sponsored opinions — just research.