Claude Opus 4.7
I’ve been meaning to do this for a while. The release cadence for frontier models is rapid (in the last thirty days alone we’ve had Mythos Preview tease, Gemini 3.1 Pro patches, a GPT-5.4 point update, and now this), and keeping up is starting to feel like a job. So I’m making it one. Every time a major model drops, I’ll write one of these.
Welcome to Model Drop. First up: Claude Opus 4.7.
Model: Claude Opus 4.7 (claude-opus-4-7)
Ship date: April 16, 2026
Maker: Anthropic
Context window: 1M tokens
Pricing: $5 / $25 per million input / output tokens (same as 4.6)
Multimodal: Text + images up to 2,576px on the long edge (~3.75MP, 3x prior Claude)
Available on: Claude.ai, Claude Code, Cursor, the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry
Headline benchmarks: 87.6% SWE-bench Verified (up from 80.8%), 64.3% SWE-bench Pro (up from 53.4%), 70% CursorBench (up from 58%), 98.5% on XBOW’s visual-acuity benchmark (up from 54.5%)
Safety class: ASL-3 class, below Mythos Preview, cyber capabilities deliberately dialed back with automated request filtering
Full system card: 232 pages (Anthropic PDF)
What shipped
Anthropic’s own framing: Opus 4.7 is a targeted upgrade over Opus 4.6 focused on hard software engineering and long-running agentic work. Their launch post leans hard into one claim, and it’s the only one that matters: you can now hand off your hardest coding work, including the messy multi-hour stuff that used to need a human babysitter, and expect it to come back done.
The early-access quotes in the launch post read like a greatest-hits album of engineering tools: Linear, Cursor, Vercel, Replit, Notion, Warp, Devin, Factory, Bolt, Hex, CodeRabbit, Ramp. Linear reported a 13% lift in resolution rate. Cursor jumped from 58% to 70% on CursorBench. Rakuten resolved 3x more production tasks. Devin said the model “works coherently for hours” on problems previous Opus models would quit on. Notion flagged it as the first model to pass their implicit-need tests. Vercel said it runs proofs on systems code before starting work, which is genuinely new behavior.
Opus 4.6 was already a mainstay in my Claude Code and Cursor setups. Opus 4.7 closes the gap on some of the failure modes I was still hitting: giving up on hard debugging, looping on tool errors, and producing work that technically runs but quietly skips the ambiguous part of the spec.
What’s new in the architecture
Anthropic doesn’t publish model architecture details, so take this as “observed behavior” rather than “confirmed internals.” A few things are new and worth knowing: