Anthropic just launched Claude Opus 4.7, and this is not being framed as a small tune-up.

It is now Anthropic’s most capable generally available model, with the company claiming meaningful gains over Opus 4.6 in advanced software engineering, long-running task execution, instruction following, vision, and professional knowledge work. Anthropic says users can hand off harder coding tasks with less supervision, and that the model is better at checking its own work before reporting back.

That is the real story here.

This is Anthropic trying to push a lot more frontier capability into the public flagship, while still keeping Mythos-class risk in a separate lane. Anthropic explicitly says Opus 4.7 is less broadly capable than Mythos Preview, but still beats Opus 4.6 across a range of important benchmarks.

There are basically two Claude stories now.

One is the highly restricted, higher-risk tier represented by Mythos Preview. The other is the model Anthropic is comfortable putting into broader use across Claude, the API, Bedrock, Vertex AI, and Microsoft Foundry. Opus 4.7 is that second lane. Anthropic says it is available today across all of those surfaces, with the same pricing as Opus 4.6 at $5 per million input tokens and $25 per million output tokens.

That matters because frontier labs are no longer releasing every major capability jump the same way.

Some models are becoming products. Others are becoming controlled infrastructure. Opus 4.7 is Anthropic trying to deliver more of the good stuff now, while still holding back the more sensitive cyber-heavy edge of the frontier. Anthropic even says it experimented during training with ways to specifically reduce the model’s cyber capabilities relative to Mythos.

This is where the launch gets more interesting.

Anthropic says Opus 4.7 improved resolution on its 93-task coding benchmark by 13% over Opus 4.6, including four tasks that neither Opus 4.6 nor Sonnet 4.6 could solve. It also says the model has faster median latency, stronger instruction following, and more rigorous behavior on complex, long-running coding work.

Outside Anthropic’s own claims, the customer evals they published are also worth paying attention to.

Cursor says Opus 4.7 hit 70% on CursorBench, up from 58% for Opus 4.6.

Harvey says it scored 90.9% on BigLaw Bench at high effort, with better reasoning calibration and stronger handling of ambiguous legal editing tasks.

Applied AI says it tied for top overall score on its six-module research-agent benchmark at 0.715, and improved on the General Finance module from 0.767 to 0.813 versus Opus 4.6.

XBOW says Opus 4.7 hit 98.5% on its visual-acuity benchmark, versus 54.5% for Opus 4.6, which is a huge jump and one of the clearest signs that Anthropic’s higher-resolution vision claims are not just marketing.

Notion says Opus 4.7 delivered 14% better performance over Opus 4.6 on complex multi-step workflows, while using fewer tokens and making about one-third the tool errors.

Rakuten says it resolves 3x more production tasks than Opus 4.6 on Rakuten-SWE-Bench.

Those are not tiny deltas.

They point to the same pattern across a bunch of real-world evals: stronger autonomy, better follow-through, fewer tool mistakes, and more reliable long-horizon execution.

This launch matters because it shows where the frontier is actually moving.

The old story was mostly about bigger benchmark numbers and raw model intelligence. The new story is about whether the model can stay coherent over long runs, use tools properly, recover from failures, read dense visual inputs, and keep going without falling apart halfway through.

That is exactly the profile Anthropic is emphasizing here.

They are not selling Opus 4.7 as a chat model with slightly better vibes. They are selling it as the practical flagship for serious coding, serious agents, and serious professional work. Their launch page keeps coming back to the same traits: rigor, consistency, self-verification, instruction following, long context performance, and higher-quality outputs across interfaces, slides, and docs.

That is a bigger shift than it looks.

It means the benchmark race is starting to matter less in isolation. What matters more is whether the model actually holds up inside real systems.

The vibe is vibing.

Anthropic is trying to commercialize more frontier capability without fully opening the door on Mythos-level risk.

That is why Opus 4.7 feels important. It gives enterprises and developers a stronger Claude right now, while letting Anthropic keep its most sensitive capabilities behind more safeguards. Anthropic says Opus 4.7 is the first model where it is testing cyber safeguards intended to eventually support broader Mythos-class releases, and it is inviting legitimate security professionals into its new Cyber Verification Program for approved use cases like penetration testing, vulnerability research, and red teaming.

So this is not just a better model.

It is Anthropic building a public flagship that is stronger, more deployable, and clearly designed to be the bridge between today’s Claude products and whatever comes next.

Anthropic just launched Claude Opus 4.7, its strongest generally available model yet, and the benchmark story is actually meaningful this time. Anthropic says it improved 13% on its 93-task coding benchmark, Cursor says it jumped from 58% to 70% on CursorBench, Harvey says it hit 90.9% on BigLaw Bench, XBOW says it leapt from 54.5% to 98.5% on visual acuity, and Notion says it improved 14% on multi-step workflows while cutting tool errors sharply.

The bigger takeaway is that Anthropic is now clearly splitting the frontier into two layers: a broader public flagship in Opus 4.7, and a more restricted high-risk tier above it. That tells you a lot about where model releases are heading next. More capability is coming, but not all of it is going to be distributed the same way.

Anthropic Claude Opus 4.7 Drop 👉 Here

Stay building. 🤖