Anthropic shipped Claude Opus 4.7 on April 16, the first generally-available release since Opus 4.6 and the production-tier sibling of the restricted Mythos Preview that drove cybersecurity headlines earlier this month. It landed simultaneously on Claude Platform, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry, which is the cleanest cross-cloud launch Anthropic has done to date. VentureBeat framed it as a narrow retaking of the "most powerful generally-available LLM" title.
The benchmark deltas are the part worth paying attention to. SWE-bench Verified moved to 87.6%, up from 80.8% on Opus 4.6. SWE-bench Pro, the harder variant, lands at 64.3%. CursorBench hit 70%, up from 58%. GPQA Diamond reached 94.2%. The feature set matters more than the numbers though. A new "xhigh" effort level exposes more compute per turn. Task budgets let you cap how much work a single task spends. /ultrareview is a multi-agent code review feature, which continues the subagent-primitive convergence we covered when Gemini CLI shipped subagents last week. Vision input tops out at 2,576 pixels on the long side (roughly 3.75 megapixels), making Opus 4.7 the first Claude model with high-resolution image support. The 1M-token context and 128k output ceiling are preserved from 4.6. Pricing holds at $5 input and $25 output per million tokens.
Two things stand out. First, SWE-bench Verified at 87.6% is a seven-point jump over Opus 4.6 in a single cycle, which is a larger swing than most recent Claude versions have produced. The CursorBench delta (58 to 70) is even bigger in relative terms. Anthropic's agentic-coding bet continues to be where the product gets the most investment, and the /ultrareview multi-agent review feature is direct evidence. Second, the simultaneous launch across Bedrock, Vertex, and Microsoft Foundry signals that Anthropic's deep-AWS partnership plus available-on-all-major-clouds posture is not the cloud-exclusivity some read into the $25B Amazon expansion. Claude remains genuinely multi-cloud at the product level, even if the spend concentration is one-way.
If you run coding agents, the practical implication is straightforward. Opus 4.7 is a drop-in swap on every major API surface at the same price point as 4.6. The /ultrareview feature is the one to experiment with first — multi-agent code review built into the model makes the subagent pattern a default rather than a configuration step. The 2,576-pixel vision upgrade unlocks screenshot-debugging workflows that were previously limited by downscaling. If you are in a regulated environment that prefers Bedrock or Vertex, parity at launch means you can adopt 4.7 with the same cloud-native posture you already use. The model that produced Mythos-tier cybersecurity results in research-preview is now in production, one notch below, generally available.