Claude Opus 4.7

Supported Functionality

Item	Specification
Input	Text, Image (up to 2576px / 3.75MP)
Output	Text
Context	1,000,000 tokens
Max Output	128,000 tokens
Vision	✓ Supported (high-resolution)
Function Calling	✓ Supported

Description

Claude Opus 4.7 is Anthropic's flagship generally available large language model, released on April 16, 2026 with model ID claude-opus-4-7, as the latest member of the Claude 4 family. It is the most capable Claude model open to all users (the stronger Claude Mythos Preview is restricted to Project Glasswing partners), purpose-built for long-horizon agentic workflows, production-grade software engineering, and complex multi-step tasks.

Compared to Opus 4.6, version 4.7 lifts SWE-bench Verified from 80.8% to 87.6% and SWE-bench Pro from 53.4% to 64.3%, while roughly tripling vision resolution (from 1.15MP to 3.75MP). The release introduces an xhigh effort tier, task budgets, adaptive thinking (replacing fixed extended-thinking budgets), and a new tokenizer — paired with Project Glasswing cybersecurity safeguards that pave the way for a broader future rollout of Mythos-class models.

Key Capabilities

• Coding: 87.6% on SWE-bench Verified, 64.3% on SWE-bench Pro, and 69.4% on Terminal-Bench — handles cross-file context in 100k-line codebases for refactoring, dependency upgrades, and batch fixes.
• Long-horizon Agents: ~14% improvement in multi-step agentic reasoning with roughly a third of the tool-call errors, capable of running autonomous workflows (CI/CD, async tasks) for hours.
• High-Resolution Vision: First Claude model to support 2576px / 3.75MP images; MathVista rises from 69.8% to 79.3%, enabling interpretation of chemical structures, engineering diagrams, and dense charts.
• Adaptive Thinking: Dynamically allocates reasoning depth based on task complexity, paired with four low/medium/high/xhigh effort levels to trade off capability against speed and cost.
• Long Context (1M): A 1-million-token context window plus 128k max output lets the model ingest entire repositories, long technical documents, or full quarterly filings in one pass.
• Instruction Following: Strictly follows instructions without silently generalizing across items or inferring requests that were never made — ideal for procedural enterprise workflows.
• Computer Use: Scores 78.0% on OSWorld-Verified, delivering state-of-the-art GUI operation across browser and desktop environments.

Technical Strengths

Feature	Benefit
Adaptive thinking replaces fixed budgets	Model decides reasoning depth autonomously, eliminating manual tuning while outperforming the older extended-thinking mode in internal evals
Task budgets	Set a total token budget across an entire agentic loop (thinking + tool calls + output), preventing runaway costs on long-running tasks
New tokenizer	Improves token efficiency for multilingual and code content and underpins capability gains (though same text may use 1.0–1.35× more tokens)
Zero operator access on Bedrock	Customer prompts and responses are invisible to Anthropic and AWS operators, meeting strict enterprise data privacy requirements
Built-in cybersecurity safeguards	Automatically detects and blocks high-risk cyber-attack requests, lowering compliance risk for enterprise deployments
Native availability across major clouds	Ships day-one on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry for streamlined enterprise adoption

Capability Ratings

Dimension	Rating	Notes
Reasoning	Top-tier	94.2% on GPQA Diamond and class-leading multi-step agentic reasoning
Coding	Top-tier	87.6% on SWE-bench Verified, among the strongest generally available coding models
Creative Writing	Excellent	Greater taste and consistency for professional writing, slides, and documents
Multimodal	Excellent	First high-resolution vision support; reads chemical structures and complex diagrams
Response Speed	Moderate	Adaptive thinking adds silent reasoning latency, though low-effort 4.7 ≈ medium-effort 4.6
Context Window	Huge	1M-token context handles full codebases and lengthy documents in a single call

Use Cases

• Production-grade software engineering: Cross-file refactors, dependency upgrades, and bug fixes that can be executed autonomously inside large monorepos.
• Long-running agentic tasks: CI/CD automation, background coding jobs (Claude Code background mode), and multi-step business workflows spanning hours.
• Enterprise knowledge work: Financial report analysis, contract review, cross-document synthesis, and decision support across full enterprise corpora.
• Professional vision tasks: Life-sciences patent workflows (e.g., Solve Intelligence), medical imaging assistance, engineering drawings, and technical documentation analysis.
• Code review and quality assurance: The /ultrareview command in Claude Code and integrations with tools like CodeRabbit deliver materially higher recall on complex PRs.
• Data-analysis agents: Serves as the analysis brain inside BI platforms like Hex, honestly reporting missing data instead of fabricating plausible fallbacks.
• Finance and legal analysis: Leading scores on GDPVal-AA and similar professional evals make it well-suited to financial modeling, regulatory document handling, and legal research.