Claude Opus 4.7 Lands — A New Baseline for Serious AI Coding

The newest member of the Claude family — Opus 4.7 — is officially out of preview and generally available across Anthropic's ecosystem. On paper it's a point release. In practice, early testers are describing it as the first model they genuinely trust to run long, complicated engineering tasks without a babysitter.

What's actually new

Opus 4.7 is the biggest jump Anthropic has shipped on the Opus line since 4.x launched. The headlines:

• Long-horizon coding that holds together. Users report they can hand Opus 4.7 the kind of multi-hour, multi-file software tasks that used to require constant human correction. It sticks to the brief, self-verifies its output, and — critically — doesn't "drift" halfway through a run.
• A genuinely upgraded vision pipeline. The model now accepts images up to 2,576 pixels on the longest edge (roughly 3.75 megapixels), more than triple prior limits. That's big for anything pixel-level: UI screenshots, dense dashboards, architecture diagrams, design references.
• Better taste. Interfaces, slide decks and documents it produces look less "AI output" and more like work a competent designer or analyst would hand you.
• Stronger on finance. Opus 4.7 leads the third-party GDPval-AA benchmark — which scores economically valuable knowledge work across finance, legal and adjacent domains — and comfortably outperforms 4.6 on rigorous financial modelling and executive-quality presentations.

It still doesn't match the raw ceiling of the restricted Mythos Preview model, but across standard benchmarks it is the clear new state-of-the-art in the generally-available tier.

Cybersecurity: the first deployment of Glasswing safeguards

Following the recent Project Glasswing disclosure — which laid out how advanced models cut both ways in cybersecurity — Anthropic is being deliberate. Mythos remains locked down precisely so its cyber protections can be tested. Opus 4.7 is the first general-availability model shipping with those protections turned on.

That means two things in practice:

1. Opus 4.7's native cyber capabilities are intentionally dialed below Mythos.
2. It ships with automated safeguards designed to detect and block prohibited or high-risk cybersecurity requests.

For legitimate defensive work — red-teaming, pentesting, vulnerability research — Anthropic has opened a dedicated Cyber Verification Program that grants verified professionals the access they need without tripping the guardrails.

Availability and pricing

Opus 4.7 is live today across:

• claude.ai (all paid tiers)
• The Anthropic API
• Amazon Bedrock
• Google Cloud Vertex AI
• Microsoft Foundry

Pricing is unchanged from the 4.6 era:

• $5 per million input tokens
• $25 per million output tokens

The API identifier is claude-opus-4-7.

What early testers noticed

A few consistent themes came out of the early-access program — some of them are real gotchas.

It takes your prompts literally

Opus 4.7 follows instructions much more strictly than prior versions. That sounds like a pure win, and on clean prompts it is. But prompts written for older, more "forgiving" models — the kind that quietly reinterpreted vague instructions or skipped parts that didn't quite fit — can now produce unexpected results. If you have a prompt library built up against 4.6, assume you'll need a refinement pass before pointing it at 4.7.

Vision unlocks new workflows

With the bigger image budget, Opus 4.7 is usable for tasks previous Claude generations simply couldn't do well: pulling numbers out of intricate diagrams, reading dense UI screenshots, referencing pixel-perfect design specs. If your pipeline used to downscale screenshots before sending them in, you can probably stop doing that now.

File-system memory is meaningfully better

Across extended, multi-session agentic workflows, the model is much better at using file-based notes as persistent memory. You'll spend far less time re-feeding it context at the start of every task.

Business and finance work

Beyond the GDPval-AA result, early financial-services users are specifically calling out its model-building and presentation quality. This is the first release where "use Claude as your analyst" stops being a stretch claim.

Safety and alignment

Alignment assessments put Opus 4.7 on roughly the same footing as 4.6 — which is to say, low rates of concerning behaviour like deception or sycophancy. Prompt-injection resistance is meaningfully improved. The one consistent note from reviewers is that it can occasionally be overly cautious — for example, padding harm-reduction advice on controlled substances with more detail than strictly needed. Overall the evaluations land in the "highly trustworthy, well-aligned" zone.

New platform features launching alongside the model

Anthropic used the launch to ship a handful of platform updates that are arguably as interesting as the model:

• xhigh effort level. A new "extra high" setting slots in between high and max, giving finer control over the reasoning-quality vs. latency tradeoff on hard problems.
• Task Budgets (API, public beta). Lets developers cap token expenditure per task and helps the model prioritize effort during long background runs.
• Claude Code upgrades.
  • A new /ultrareview command that does a deep-dive code review aimed at the kinds of subtle bugs and design flaws a careful human reviewer would catch.
  • Expanded auto mode for Max users — the model can make more autonomous decisions to avoid interrupting the human during long coding sessions.

Migrating from 4.6 to 4.7 — two things to watch

Upgrading is mostly a drop-in, but two factors can quietly increase your bill if you're not paying attention:

1. New tokenizer. The same input text can map to roughly 1.0–1.35x the tokens of 4.6, depending on content. Assume input cost creeps up a little even before you change anything else.
2. Longer "thinking" at high effort. At upper effort levels, particularly in agentic setups, Opus 4.7 simply thinks longer — which produces more output tokens. You get better answers on hard problems, but the meter runs.

Practical advice for anyone rolling it out in production:

• Monitor per-task token usage for the first week after switching.
• Tune the effort parameter to the task — you don't need xhigh for everything.
• Use Task Budgets where long agent runs are involved.
• Where latency or cost matters, explicitly prompt for concise output. Opus 4.7 responds well to that instruction.

Bottom line

Opus 4.7 isn't a headline-grabbing reinvention. It's the opposite — a release where the chart-topping benchmark numbers are almost the least interesting part. The real story is that a generally-available model can now be pointed at a long, messy engineering or analytical task and come back with something worth shipping. That's the capability bar most teams have actually been waiting for.

If you're already on 4.6, the upgrade is worth doing — just plan for a short prompt-revision pass, and keep an eye on your token usage for the first few days.