In-depth Introduction to Claude Opus 4.7: Coding Capabilities Upgraded, 1M Context at No Extra Cost, What Are the Real-World Drawbacks?

📄Full Article· Automatically extracted by trafilaturaGemini 翻譯4825 words

Anthropic released Claude Opus 4.7 on April 16, featuring a significant leap in coding capabilities (SWE-bench Pro hit 64.3%, an 10.9-point improvement over the previous 4.6 version's 53.4%), a 3x increase in visual resolution, and a 1M context window at standard pricing without a long-context premium. However, some users have reported several pain points. (Previous context: Has Claude Code become dumber? AMD AI Director fires back: Reasoning depth plummeted by 67%, API monthly costs surged 144x) (Background: The Big Short's Michael Burry claims: Anthropic is eating Palantir) On April 16, Anthropic officially pushed Claude Opus 4.7 to four major cloud platforms: API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. GitHub Copilot Pro+, Business, and Enterprise users also received the update simultaneously. Compared to version 4.6, the direction of this upgrade is clear: strengthening coding capabilities is the absolute main focus, vision is the second priority, and API changes are the most noticeable and controversial parts for long-time users. In this article, I have summarized the new features, my personal testing experience, and the polarized feedback from the community regarding its drawbacks. To understand Opus 4.7, one must first understand its position in Anthropic's model hierarchy. Internally, Anthropic has a higher-capability model called Mythos. Due to its ability to discover zero-day vulnerabilities at scale, it is deemed a national security risk and is currently not released publicly, reserved only for federal agencies and key infrastructure vendors on the Project Glasswing partnership list. Opus 4.7 is the "strongest publicly available" version for civilian use. This is the underlying strategy of Anthropic. CEO Dario Amodei estimated in a Financial Times interview that Chinese open-source models will catch up to Mythos's capabilities in 6 to 12 months at the earliest. Until then, Opus 4.7 is the strongest commercial model Anthropic can provide to the market. Among the new feature list, these three upgrade points are the most impressive: 1. Significant leap in coding capabilities. SWE-bench Pro (a benchmark evaluating AI's ability to solve real GitHub issues) jumped from 53.4% in 4.6 to 64.3%, an improvement of 10.9 points. This score ranks first among public models, surpassing GPT-5.4's 57.7% and Gemini 3.1 Pro's 54.2%. SWE-bench Verified also rose from 80.8% to 87.6%. CursorBench, which specifically measures the autonomous coding performance of Cursor IDE, jumped from 58% to 70%. Anthropic's own claim is: "Coding tasks that you previously needed to supervise closely before submitting can now be entrusted to 4.7 with more confidence." 2. Multi-step agentic workflow (allowing Claude to complete tasks involving multiple tool calls at once) feels very responsive. The official claim is a 14% improvement over 4.6, using fewer tokens, with tool usage errors reduced to about 1/3. This figure aligns well with my own experience, which I will discuss in the next section. 3. 3x improvement in visual capabilities. Opus 4.7 is the first Claude model to support high-resolution images, up to 2576px / 3.75MP, which is three times that of 4.6. In practice, you can drop screenshots directly without resizing. For PDF charts, IDE screenshots, design drafts, etc., it can retain more details. In terms of benchmarks, visual-related benchmarks saw a 13% improvement. 1M token context without price hike. The fact that the context window expanded from 200K to 1M without a price increase is a point of great concern. Users were previously worried about price hikes or having the context window shrunk back to 200K. More importantly, Anthropic did not charge a long-context premium. Previously, Claude requests exceeding 200K tokens incurred extra costs; 4.7 now follows the standard pricing of $5 / $25 per million tokens across the board. Combined with the 128K max output tokens, workflows like long codebase analysis, long contract reviews, and long log diagnostics no longer require splitting into chunks. This area saw the most changes in the new feature list and sparked much controversy. Added xhigh (extra high) effort level. Previously, Claude's reasoning power had two levels: high and max. 4.7 inserted

Data Status✓ Full text extractedRead Original (動區 BlockTempo)

🔍Historical Similar Events· Keyword + Asset Matching6 items

2026-04-23

OpenAI Launches GPT-5.5 to Challenge Anthropic’s Claude Opus 4.7

Similarity 130%關鍵字 claude/opus

2026-04-21

This Frankenstein AI Merges Claude Opus, GLM and Qwen—And Outperforms Top Models

Similarity 130%關鍵字 claude/opus

2026-04-17

Anthropic Debuts Claude Opus 4.7 as Agentic Workflows Take Center Stage

Similarity 130%關鍵字 claude/opus

2026-04-16