Claude Opus 4.7 Explained

Claude Opus 4.7 is Anthropic's advanced large language model released April 16, 2026, designed for complex software engineering, long-horizon agentic tasks, and multimodal understanding with significantly improved vision capabilities (up to 2,576 pixels on the long edge). Opus 4.7 costs $5 per million input tokens and $25 per million output tokens and runs on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.

The model introduces a new tokenizer that produces roughly 30% more tokens for the same text compared to earlier Claude models, which affects real-world pricing despite the unchanged per-token rate. Opus 4.7 includes the new xhigh effort setting for deeper reasoning, task budgets to control agent spending, and improved instruction-following that executes prompts more literally than prior versions.

On coding benchmarks, Opus 4.7 jumped to 87.6% on SWE-bench Verified (up from 80.8% on Opus 4.6), earning a 64.3% score on SWE-bench Pro—the highest among publicly available non-preview models. It also delivers state-of-the-art performance on legal reasoning (90.9% on BigLaw Bench), agentic workflows, and finance document analysis.

Opus 4.7 represents an intermediate step in Anthropic's lineup—its predecessor is Opus 4.6, and its successor is Opus 4.8 (released May 28, 2026). It is a highly capable Opus-tier model for complex reasoning tasks, though Opus 4.8 (released May 2026) is now Anthropic's most capable Opus-tier model, and Claude Mythos Preview (launched April 2026 via Project Glasswing) and Claude Mythos 5 (released June 2026) represent a tier above Opus.

What Is Claude Opus 4.7?

Claude Opus 4.7 is Anthropic's advanced language model for complex reasoning, released April 16, 2026 as a successor to Opus 4.6. Developed by Anthropic, a San Francisco-based AI safety company, the model runs on a 1 million token context window and supports text and image input. Weights are proprietary and not available open-source; the model is accessed via API or third-party platforms only.

Anthropic uses the model ID 'claude-opus-4-7' on the Claude API, 'anthropic.claude-opus-4-7' on AWS Bedrock, and 'claude-opus-4-7' on Google Cloud Vertex AI. It was released alongside Opus 4.8 (released May 28, 2026) and comes with a new tokenizer and xhigh effort level for graduated reasoning control.

The model's training data includes knowledge through January 2026. Anthropic notes in its announcement that Opus 4.7 is 'largely well-aligned and trustworthy' with low rates of concerning behavior, though it has modest weaknesses in harm-reduction advice on controlled substances.

Key Features & Capabilities

Opus 4.7 excels at vision processing with support for images up to 2,576 pixels on the long edge—approximately 3.75 megapixels and over 3x the resolution of prior Claude models. The model can interpret dense technical diagrams, chemistry structures, professional screenshots, and multimodal documents with strong accuracy.

Advanced coding and software engineering is a core strength. Opus 4.7 can reason about complex GitHub issues, refactor large codebases, and handle long-horizon agentic tasks with 'particular gains on the most difficult tasks.' It supports tool use for autonomous problem-solving with improved error handling.

The xhigh effort level (new in Opus 4.7) enables deeper chain-of-thought reasoning without exposing thinking blocks or adding latency overhead. For enterprise knowledge work, the model delivers stronger performance on financial documents, legal files, and multi-step research workflows.

Task budgets (in public beta) let developers set token spend caps per autonomous agent task. Prompt caching reduces input costs to 10% of standard rate on repeated context (after a 1.25x write cost). The model follows instructions literally, executing the exact text provided rather than interpreting intent loosely.

Benchmarks & Performance Scores

On coding tasks, Opus 4.7 leads all public non-preview models: 87.6% on SWE-bench Verified (software engineering: 80.8% on 4.6), 64.3% on SWE-bench Pro (resolving real GitHub issues, up from 53.4%), and 70% on CursorBench versus 58% for Opus 4.6.

For reasoning and knowledge, the model scores 94.2% on GPQA Diamond (graduate-level physics, chemistry, biology reasoning). On legal analysis, it reaches 90.9% on BigLaw Bench. On agentic tool use, it scores 77.3% on MCP-Atlas.

Vision capability improved dramatically: 98.5% on XBOW visual acuity (versus 54.5% for Opus 4.6). Document understanding shows 21% fewer errors than Opus 4.6 on OfficeQA Pro. Finance agent tasks deliver state-of-the-art performance; GitHub Copilot adoption in April 2026 signals production readiness.

A notable weakness is web research: BrowseComp dropped to 79.3% from 83.7%, trailing GPT-5.4 Pro (89.3%). Opus 4.7 is not the right choice if multi-page synthesis and fact-checking from live web sources are critical.

Pricing & Cost Structure

Standard API pricing is $5 per million input tokens and $25 per million output tokens (unchanged from Opus 4.6). With prompt caching, cache writes cost 1.25x base rate (5-minute duration) or 2x (1-hour duration); cache reads cost 10% of base input rate. Batch API processing costs $2.50 input / $12.50 output—a 50% discount.

The critical operational detail: Opus 4.7's new tokenizer produces approximately 30% more tokens for identical text versus Opus 4.6. A prompt that cost $0.50 on Opus 4.6 may cost $0.65 on Opus 4.7 (30% token inflation), even though per-token rates stayed flat. Budget for 0–35% cost increase per request depending on content type.

Fast mode (in research preview, deprecated July 24, 2026) charged $30 input / $150 output but is no longer recommended. Prompt caching and batch processing are the primary levers for cost control. The model is available on Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.

For high-volume users, contact Anthropic sales for volume discounts and custom pricing. US-only inference costs 1.1x the standard rate on all token types.

Where Claude Opus 4.7 Excels

Production code review and long-horizon engineering tasks are Opus 4.7's core use cases. Teams use it for resolving complex GitHub issues, shipping production code, and autonomous coding agents. It produces more careful output than faster alternatives and handles edge cases better.

Enterprise knowledge work including financial document analysis (investment memos, 10-Ks, quarterly earnings transcripts), legal briefs, and multi-step research. Hebbia's orchestrator and Databricks' Genie leverage Opus models for finance and data workflows where citation accuracy and token efficiency matter.

Agentic AI systems that coordinate multiple tool calls (API calls, database queries, web searches) benefit from Opus 4.7's improved tool-use reliability and 'one-third fewer tool errors' versus Opus 4.6. Task budgets let you set spend caps per autonomous agent run.

Multimodal document understanding and dense visual reasoning—technical diagrams, chemistry structures, professional UI screenshots, and dense PDFs. The 3.75 megapixel vision input enables interpretation of information-dense materials where lower-resolution models fail.

Limitations & Drawbacks

Web research and multi-page synthesis is a weakness. BrowseComp dropped 4.4 points (83.7% to 79.3%) versus Opus 4.6 and trails GPT-5.4 Pro (89.3%). If your workflow requires fact-checking across multiple sources or synthesizing recent web information, benchmark Gemini 3.1 Pro or GPT-5.5 first.

Token inflation from the new tokenizer creates cost uncertainty. Existing prompts and rate-limit calculations may consume 30% more tokens than expected. Systems that rely on exact token counts for budget forecasting or rate-limit planning will need recalibration.

Strict literal instruction-following can break workflows that rely on interpretive flexibility. The model 'executes the exact text provided'—which is more correct but may conflict with prompts written for older models that interpreted intent loosely. Reword prompts to be explicit.

Deliberate cybersecurity capability reduction. Anthropic intentionally capped cybersecurity performance (73.1% on CyberGym) by applying 'differential capability reduction during training.' The released Claude Mythos Preview scores 83.1% on the same benchmark, suggesting Opus 4.7 is not the right choice for offensive security work.

Slower than Sonnet 5. Opus 4.7 has 'moderate' latency compared to Sonnet 5's 'fast' latency, and it costs 1.67x more per input token. For latency-sensitive use cases or production chat applications, Sonnet 5 is better.

How Claude Opus 4.7 Compares to Alternatives

Claude Opus 4.7 vs GPT-5.5 (OpenAI, April 2026): Opus leads on coding (64.3% SWE-bench Pro vs 58.6%), reasoning (94.2% GPQA Diamond vs 93.6%), and hallucination resistance (36% vs 86% long-form factuality). GPT-5.5 is faster and makes fewer tool calls for agentic tasks. GPT-5.5 costs $5 input / $30 output and includes 1 million token context window (matching Opus 4.7's context length).

Claude Opus 4.7 vs Gemini 3.1 Pro (Google, February 2026): Gemini 3.1 Pro costs $2/$12 (under 200K tokens) and supports the same 1M context window. Opus 4.7 leads on coding and long-form accuracy; Gemini 3.1 Pro is 6x cheaper on input and excels at volume-heavy reasoning tasks. For web research, Gemini 3.1 Pro outperforms both.

Claude Opus 4.7 vs Claude Sonnet 5 (Anthropic, June 2026): Sonnet 5 costs $2/$10 through August 31, 2026 (then $3/$15), is 2–3x faster, and offers best balance of speed and intelligence for production. Opus 4.7 is reserved for the most demanding reasoning tasks where quality differentiates revenue. Sonnet 5 has extended thinking and better latency.

Claude Opus 4.7 vs Claude Opus 4.8: Opus 4.8 (released May 28, 2026) adds stronger agentic autonomy, better error-detection in code review, and improved computer-use capabilities. Opus 4.7 is cheaper and still excellent for most reasoning tasks. Opus 4.8 is Anthropic's most capable Opus-tier model as of May 2026.

How to Get Started with Claude Opus 4.7

For API developers: Sign up at claude.ai or the Claude Console, create an API key, and use model ID 'claude-opus-4-7' in your requests. See Anthropic's quickstart guide for code examples in Python, JavaScript, and cURL.

For AWS users: Access Claude Opus 4.7 via Amazon Bedrock or Claude Platform on AWS using the Bedrock ID 'anthropic.claude-opus-4-7'. Bedrock offers regional endpoints for data residency with a 10% pricing premium.

For Google Cloud: Use claude-opus-4-7 in Vertex AI with global, multi-region, or regional endpoints. Regional routing incurs a 10% premium.

For enterprise: Contact Anthropic sales (sales@anthropic.com) for volume discounts, custom token limits, dedicated support, and negotiated pricing. Microsoft Foundry users access Opus 4.7 through Claude in Microsoft Foundry.

Compliance, Safety & YMYL Considerations

For healthcare workflows: Claude is deployed in HIPAA-ready infrastructure allowing protected health information (PHI) processing without using data for model training. Opus 4.7 is suitable for healthcare decision support but should always be paired with human clinician review for treatment decisions.

For finance and legal work: Opus 4.7 reaches 90.9% accuracy on legal reasoning benchmarks and excels at financial document analysis. However, it is still an AI system—regulatory guidance requires human review on trades, loan approvals, and client-facing legal advice. Use it to accelerate document review and preliminary analysis, not to replace professional judgment.

Safety profile: Anthropic notes Opus 4.7 is 'largely well-aligned and trustworthy' with low rates of concerning behavior. The model resists prompt injection attacks and can identify harmful requests even when reframed across successive messages. It has modest weaknesses in harm-reduction advice on controlled substances.

Metadata and citations: Unlike some competitors, Opus 4.7 does not include inline citations by default. For regulated use, implement citation tracking in your application layer. Prompt caching helps reduce costs on repeated compliance documents.