Anthropic Just Released Opus 4.7. I've Been Using It All Day and the Difference Is Not Subtle.

Opus 4.6 came out on February 5th. Opus 4.7 dropped today, April 16th. That's ten weeks. Ten weeks between major model releases from the same company.

Let me put that in perspective. In 2024, Anthropic released maybe two or three models the entire year. In 2026, they've already shipped Opus 4.6, Sonnet 4.6, Opus 4.7, and Mythos Preview — four models in under four months. If they keep this pace, we're looking at a new frontier model roughly every two and a half months. That's not normal. That's not how this industry used to work.

And the jumps aren't incremental. Opus 4.7 scored 87.6% on SWE-bench Verified, up from 80.8% on Opus 4.6. That's a seven-point leap in ten weeks. It leads GPT-5.4 and Gemini 3.1 Pro on both coding benchmarks. It resolves three times more production tasks than its predecessor. And it's the first Claude model that can figure out what tools to use without being explicitly told — Anthropic calls them "implicit-need tests."

But I don't need benchmarks to tell me it's better. I've been using it all day.

I'm building a search agent — something that can research topics at the depth and quality of Perplexity. It's the kind of project that requires planning out multiple systems: web scraping, result ranking, source verification, answer synthesis. The kind of build where the AI needs to think architecturally, not just spit out code snippets.

Opus 4.7 planned this build out at a level of detail I haven't seen before. I'm talking about a full system architecture with clear separation of concerns, edge case handling, and implementation steps that actually made sense when I read through them. It didn't just give me code — it gave me a blueprint. And then when I asked it to build each piece, the code matched the plan.

That's the difference I'm noticing. It's not just that the code is better. It's that the thinking behind the code is better. The model is planning further ahead, catching dependencies I wouldn't have thought of, and producing work that holds together as a coherent system instead of a pile of functions that sort of work.

The vision upgrade is real too. It now handles images up to 3.75 megapixels, which is over three times the previous limit. If you're working with documents, diagrams, or screenshots, that matters. And the new multi-session memory means it actually remembers context across long work sessions instead of starting from scratch every time.

Here's what I keep coming back to, though. This pace.

If Anthropic can deliver a seven-point SWE-bench improvement every ten weeks, where does that put them by the end of the year? Two more releases? Three? And each one stacking on top of the last? The compounding effect of that kind of iteration speed is hard to overstate.

They also admitted something interesting — Opus 4.7 is deliberately less capable than Mythos in cybersecurity. They baked in safeguards that reduce its ability to find and exploit vulnerabilities. Mythos stays locked behind Project Glasswing for the 12 partner organizations. Opus 4.7 is what Anthropic considers safe to release publicly. The gap between what they can build and what they're willing to ship is something worth paying attention to.

I said in an earlier post that I was bothered by how Anthropic presents their revenue numbers. That's still true. But when they ship a model this good, this fast, at the same price as the last one? It's hard to stay annoyed for long.

Sources: Anthropic, CNBC, The Next Web, 9to5Mac, GitHub Changelog