Claude Opus 4.7 vs. Opus 4.6: What Has Actually Improved?

Anthropic released Claude Opus 4.7 on April 16, 2026, about two months after Opus 4.6. It's a same-tier upgrade, not a new model family, and it works as a drop-in replacement through the same API. This guide covers what improved, what got worse, and who should switch.

One note up front: Anthropic has released newer models since, including Claude Opus 5.5, so this comparison matters most if you run workloads on Opus 4.6 or 4.7 today.

Quick Comparison

• Coding: Clearly better in 4.7
• Vision: Much better, with over 3x the image resolution
• Agent reliability: Better, with self-checking and fewer loops
• Web research: Slightly worse in 4.7
• Price per token: Identical, at $5 per million input tokens and $25 per million output tokens

Coding: The Biggest Gain

Software engineering is where Opus 4.7 pulls ahead. On SWE-bench Pro, which tests real-world coding tasks, it scores 64.3% against 53.4% for Opus 4.6. On SWE-bench Verified, it reaches 87.6% against 80.8%. Cursor reports 70% on its CursorBench against 58% for Opus 4.6.

Partner results point the same way. Rakuten says Opus 4.7 resolves three times more production tasks than Opus 4.6, with better code and test quality. Anthropic adds that it solved some tasks that Opus 4.6 couldn't.

Better at Long, Multi-Step Work

Anthropic describes this as an agent reliability release. Opus 4.7 loops less, keeps going through tool failures, and checks its own work before reporting a task done. That helps most for autonomous coding and tool-heavy workflows.

Vision: Over 3x the Resolution

Opus 4.7 accepts images up to 2,576 pixels on the long edge, about 3.75 megapixels. That's more than three times what earlier Claude models handled. One partner, XBOW, reports a visual-acuity score of 98.5% for Opus 4.7 against 54.5% for Opus 4.6. Visual reasoning on the CharXiv benchmark also rose from 69.1% to 82.1%. If you work with dense screenshots, diagrams, or documents, this is a real improvement.

New Effort Level and Features

Opus 4.7 adds an "xhigh" effort level between high and max. Anthropic says 4.7 beats 4.6 at every effort level, and low-effort 4.7 can match medium-effort 4.6 on quality. Claude Code also gains an "/ultrareview" command for code review.

Where Opus 4.7 Is Not Better

• Web research: BrowseComp fell from 83.7% to 79.3%. If your product depends on agentic web search, test before switching.
• Literal instructions: Opus 4.7 follows prompts more strictly. Loose prompts written for 4.6 may need rewriting.
• Token counts: A new tokenizer can produce 1.0 to 1.35 times more tokens for the same text.

Pricing and API

Per-token pricing stays the same: $5 per million input tokens and $25 per million output tokens. The model ID is claude-opus-4-7, and it runs on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Because of the new tokenizer, your real cost per task can shift slightly. Anthropic advises measuring it on your own traffic.

Safety

Anthropic's automated behavioral audit shows a modest improvement in Opus 4.7 on misaligned behavior compared with Opus 4.6.

Which Model Should You Use?

Choose Opus 4.7 if you run coding agents, work with large images or documents, or build long autonomous workflows. Stay on Opus 4.6 if your product relies on web browsing, or you can't retest your prompts yet. If you start a new project today, check Anthropic's documentation for the current lineup, since newer models are available.

Opus 4.7 delivers real gains in coding, vision, and agent reliability at the same per-token price. Its weak spots are web research and stricter prompt handling. Test it on your own tasks before you migrate.