Claude Opus 4.7: What's New and Why It Matters

Anthropic Just Shipped Its Most Capable Public AI Model Yet

Today, Anthropic released Claude Opus 4.7 — the latest version of their flagship AI model, and by their own account, the most powerful one they've ever made publicly available. If you use Claude for writing, research, or getting help with day-to-day tasks, this update matters. And if you don't use Claude at all, the story behind this release is interesting enough to read anyway — because buried in today's announcement is a detail that says a lot about where AI is actually headed.

Let's break it down, without any jargon.

What Is Claude Opus 4.7, Exactly?

Think of Claude as Anthropic's version of ChatGPT — an AI assistant you can talk to, ask questions, have it write things, help debug problems, or just think through complicated decisions with you. Anthropic releases new versions periodically, and each one is generally smarter, faster, or more capable than the last.

Opus is the "flagship" tier — the most powerful model Anthropic makes. Below it, there's Sonnet (the middle tier) and Haiku (the fast, lightweight option). Opus 4.7 is the latest leap forward in that top tier.

What's new in 4.7? Three things stand out.

1. It Can Now Check Its Own Work

This might sound small, but it's actually one of the most meaningful improvements in the release.

Previous AI models would confidently give you an answer — and sometimes that answer would be wrong. Not dramatically wrong, but subtly wrong. A flipped number, a misread detail, a logical leap that didn't hold up under scrutiny. The model didn't "know" it had made an error because it had no mechanism to look back.

Opus 4.7 changes this. During complex tasks — particularly coding and planning — the model now catches its own logical faults before giving you the result. Anthropic describes it as the model "accelerating through the execution phase after catching errors in the planning phase."

Translate that into everyday terms: if you ask it to help you draft a business proposal with specific numbers, it's now more likely to notice if a calculation doesn't add up before handing it over to you. It's not perfect — no AI is — but the self-correction loop is a genuine shift in reliability.

2. Its Vision Got a Massive Upgrade

If you've ever uploaded an image to an AI and asked it to describe, analyse, or extract information from it, you'll appreciate this one.

Opus 4.7 can now process images at up to 2,576 pixels on the long edge — more than three times the resolution of previous Claude models. In plain terms: it can now read fine print in a photo, understand a dense chart, or work with a detailed diagram that would have previously been too small for it to parse reliably.

This is particularly useful for people who work with visual information — designers reviewing mockups, analysts working with charts, or anyone who wants to snap a photo of a document and have the AI extract information from it accurately.

3. It Thinks Harder When It Needs To

One of the ongoing criticisms of AI models is that they sometimes apply the same "effort level" to every question — treating a simple task the same as a complex one. Opus 4.7 introduces a new reasoning mode Anthropic calls xhigh, which sits between the previous "high" and "max" settings.

The idea is straightforward: for genuinely hard problems — where accuracy matters more than speed — you can now ask the model to lean into deeper reasoning without immediately jumping to the most expensive (and slowest) compute setting. It's a finer dial for getting the balance right between "good enough fast" and "correct, even if slower."

For most everyday users this won't come up directly, but if you build things with the Claude API, or if you're using Claude for something where mistakes are costly, this new mode is worth knowing about.

The Benchmark Story: Beating ChatGPT and Gemini

Every AI release comes with a benchmarking scorecard, and Opus 4.7 lands well. Anthropic says the new model outperforms Claude Opus 4.6 (the previous version), ChatGPT 5.4, and Google Gemini 3.1 Pro across key tests for coding, reasoning, and real-world task completion.

Benchmarks aren't everything — they measure performance on curated tests, not necessarily how useful the model feels in your specific situation. But they do provide a consistent measure of progress, and Opus 4.7 clears the bar across the board.

The pricing, notably, stays the same as Opus 4.6. You get a meaningfully better model for the same cost, which is a pattern the industry has been delivering on consistently.

The Part Nobody Is Talking About Enough: Mythos

Here's where the story gets interesting.

Buried in today's announcement is a quiet acknowledgment: Opus 4.7 is not Anthropic's most powerful model. It's their most powerful publicly available model. The most capable model they've built — called Mythos Preview — hasn't been released to the public at all. It's only been shared with a small, handpicked group of technology and cybersecurity companies, as part of a private initiative Anthropic calls Project Glasswing.

This is unusual. Most AI labs release their top model to the public and keep pre-release versions internal. Anthropic is doing the opposite: holding back their frontier model entirely, letting only select organisations access it, while releasing a capable-but-deliberately-constrained version to everyone else.

Why? The short answer is safety. Mythos was trained with significantly more advanced cybersecurity capabilities — the kind that could be genuinely dangerous in the wrong hands. Anthropic made a deliberate decision during training to reduce those capabilities in Opus 4.7, building in automatic safeguards that detect and block requests involving prohibited or high-risk uses.

Opus 4.7, by design, is less capable in certain dangerous domains. That's not a flaw — it's a deliberate choice. And the fact that Anthropic openly says "yes, we have a more powerful model we're not giving you yet" is a refreshing level of transparency in an industry that rarely talks about what it's holding back.

Whether you think that's responsible AI development or unnecessary paternalism probably depends on your priors about how dangerous these tools actually are. But the signal is worth paying attention to: there's a ceiling above the ceiling.

What This Means for Everyday Users

If you're not building software or running an AI startup, here's the practical summary:

Claude just got noticeably better at the things most people use it for. Writing assistance, research, answering complex questions, working through documents — all of these will feel more accurate and reliable. The vision upgrade in particular is a quality-of-life improvement that will show up in subtle ways every time you share an image or screenshot with it.

It's available right now. Opus 4.7 is live across all of Anthropic's products — the Claude website, the mobile apps, and through Amazon, Google, and Microsoft's cloud platforms for developers. No waitlist, no phased rollout.

The price didn't change. For API users, it's $5 per million input tokens and $25 per million output tokens — same as 4.6. Prompt caching brings that down significantly for repeated contexts.

A Moment of Perspective

The pace of these releases is worth sitting with for a second.

Claude Opus 4.6 came out in February 2026. Opus 4.7 is out today, in April. Two months. Every two months, the most capable publicly available AI model gets replaced by something meaningfully better. The curve isn't flattening.

What's strange about this moment is that the model being celebrated today — Opus 4.7, which beats the best publicly available models from OpenAI and Google — is already being described by its own creators as "less broadly capable" than something they're not showing us yet.

That's either reassuring or unsettling depending on how you read it. Either the safety-first approach is working — holding back the frontier until it's been stress-tested — or the frontier is moving faster than anyone's ability to understand what's coming.

Probably both things are true at once.

For now, Opus 4.7 is here, it's good, and it's the best version of Claude you've ever been able to use. That's worth something, even in a week where the more interesting story is the one Anthropic isn't fully telling yet.

References

I use AI tools to help research and draft posts. The ideas, opinions, and takes are mine. Verify anything technical or time-sensitive before acting on it.