概要
Claude Opus 4.7, released April 16, 2026, is Anthropic's most capable broadly available model and a direct upgrade to Opus 4.6. It targets agentic software engineering, high-fidelity vision tasks, and long-running autonomous workflows. The model introduces self-verification behavior (it checks its own outputs before reporting back), literal instruction following, higher-resolution image processing (3.75MP, up from 1.15MP), and improved file-system memory for multi-session agent work. Pricing is unchanged at $5/$25 per million input/output tokens, though a new tokenizer inflates effective costs by 10–35% depending on content type.
Opus 4.7 sets new benchmarks for Anthropic on SWE-bench Verified (87.6%), SWE-bench Pro (64.3%), and MCP-Atlas (77.3%), with particularly strong gains on the hardest coding problems. Partner testimonials from Cursor, Replit, Notion, Vercel, and others confirm real-world production improvements of 10–15% in task success rates and 3× more production task resolution on some benchmarks. The model introduces a new 'xhigh' effort level between 'high' and 'max', and ships with task budgets (public beta) for cost control on long-running agent loops.
However, the release comes with notable trade-offs. Long-context multi-needle retrieval collapsed on MRCR v2 (78.3% → 32.2% at 1M tokens), BrowseComp regressed nearly 5 points, and the new tokenizer raises actual costs despite unchanged pricing. Anthropic positions Opus 4.7 below its unreleased Claude Mythos Preview, and describes it as the first broadly released model carrying cybersecurity safeguards from Project Glasswing — including automated blocking of high-risk cyber requests. The system card rates it 'largely well-aligned and trustworthy, though not fully ideal.'
ベンチマーク＆性能
## Detailed Benchmark Comparison
### Agentic Coding
| Benchmark | Opus 4.7 | Opus 4.6 | GPT-5.4 | Gemini 3.1 Pro |
|---|---|---|---|---|
| SWE-bench Verified | **87.6%** | 80.8% | 84.1% | 80.6% |
| SWE-bench Pro | **64.3%** | 53.4% | 57.7% | 54.2% |
| Terminal-Bench 2.0 | 69.4% | 65.4% | **75.1%** | — |
| CursorBench | **70%** | 58% | — | — |
| MCP-Atlas | **77.3%** | 62.7% | 68.1% | 73.9% |
### Reasoning & Knowledge
| Benchmark | Opus 4.7 | Opus 4.6 | GPT-5.4 | Gemini 3.1 Pro |
|---|---|---|---|---|
| GPQA Diamond | 94.2% | 91.3% | **94.4%** | 94.3% |
| HLE (with tools) | 54.7% | 53.1% | **58.7%** | — |
| HLE (no tools) | 46.9% | 40.0% | — | — |
| MMMLU | 91.5% | 91.1% | — | — |
### Agentic & Browser
| Benchmark | Opus 4.7 | Opus 4.6 | GPT-5.4 | Gemini 3.1 Pro |
|---|---|---|---|---|
| BrowseComp | 79.3% | 84.0% | **89.3%** | 85.9% |
| Finance Agent (SOTA) | **64.4%** | 60.7% | 61.5% | 59.7% |
| OSWorld-Verified | **78.0%** | 72.7% | — | — |
| CyberGym | 73.1% | 73.8% | — | — |
### Vision
| Benchmark | Opus 4.7 | Opus 4.6 | Delta |
|---|---|---|---|
| CharXiv-R (with tools) | **91.0%** | 77.4% | +13.6 |
| CharXiv-R (no tools) | **82.1%** | 68.7% | +13.4 |
| ScreenSpot-Pro | **79.5%** | 69.0% | +10.5 |
### Long-Context Retrieval (Major Regression)
| Benchmark | Opus 4.7 | Opus 4.6 | Delta |
|---|---|---|---|
| MRCR v2 8-needle @ 256k | 59.2% | 91.9% | **−32.7** |
| MRCR v2 8-needle @ 1M | 32.2% | 78.3% | **−46.1** |
| GraphWalks BFS (128k) | 58.6% | — | — |
### Document & Office
| Benchmark | Opus 4.7 | Opus 4.6 |
|---|---|---|
| OfficeQA Pro | **80.6%** | 57.1% |
| BigLaw Bench | **90.9%** | — |
### Safety & Alignment
| Metric | Opus 4.7 | Opus 4.6 |
|---|---|---|
| Misaligned behavior score | **2.46** (lower=better) | 2.76 |
| Self-rated sentiment (7-pt) | **4.49** | 3.74 |
| Anti-hack cheat rate | **12.5%** | 37.5% |
**Key takeaway:** Opus 4.7 dominates agentic coding and vision benchmarks, holds near-parity on knowledge/reasoning (GPQA, MMMLU), but suffers severe regressions on long-context needle retrieval and modest drops on BrowseComp. It trails GPT-5.4 on Terminal-Bench and HLE with tools.
詳細比較
## Head-to-Head Comparisons
### Claude Opus 4.7 vs GPT-5.4
| Dimension | Claude Opus 4.7 | GPT-5.4 |
|---|---|---|
| SWE-bench Verified | **87.6%** | 84.1% |
| SWE-bench Pro | **64.3%** | 57.7% |
| GPQA Diamond | 94.2% | **94.4%** |
| MCP-Atlas | **77.3%** | 68.1% |
| Terminal-Bench 2.0 | 69.4% | **75.1%** |
| BrowseComp | 79.3% | **89.3%** |
| HLE (with tools) | 54.7% | **58.7%** |
| Context Window | 1M | Not reported |
| Pricing | $5/$25 per 1M | Not reported |
**Analysis:** Opus 4.7 leads on agentic coding (+6.6 pts SWE-bench Pro) and multi-tool orchestration (+9.2 pts MCP-Atlas). GPT-5.4 is stronger on web research (+10 pts BrowseComp), terminal workflows (+5.7 pts), and hard knowledge questions (+4 pts HLE with tools). For software engineering agent teams, Opus 4.7 is the stronger pick; for research-heavy web browsing agents, GPT-5.4 retains a clear advantage. GPQA Diamond is essentially tied (~94% for both).
### Claude Opus 4.7 vs Gemini 3.1 Pro
| Dimension | Claude Opus 4.7 | Gemini 3.1 Pro |
|---|---|---|
| SWE-bench Verified | **87.6%** | 80.6% |
| SWE-bench Pro | **64.3%** | 54.2% |
| GPQA Diamond | 94.2% | 94.3% |
| MCP-Atlas | **77.3%** | 73.9% |
| BrowseComp | 79.3% | **85.9%** |
| Finance Agent | **64.4%** | 59.7% |
| Context Window | 1M | Not reported |
| Pricing | $5/$25 per 1M | Not reported |
**Analysis:** Opus 4.7 leads across most coding and agentic benchmarks. The SWE-bench Pro gap is 10.1 points — substantial for production workflows. GPQA Diamond is a virtual tie. Gemini 3.1 Pro still leads on BrowseComp (+6.6 pts), suggesting better web research synthesis. For financial analysis agents, Opus 4.7's 64.4% Finance Agent score is state-of-the-art at launch.
### Claude Opus 4.7 vs Opus 4.6 (Predecessor)
The 4.7 → 4.6 comparison reveals clear trade-offs:
- **Gains:** SWE-bench Pro +10.9, MCP-Atlas +14.6, OfficeQA Pro +23.5, visual reasoning +13 pts, file-system memory
- **Regressions:** MRCR v2 @ 1M −46.1 pts, BrowseComp −4.7 pts, new tokenizer adds 10–35% cost
- **Behavioral shifts:** More literal instruction following, more direct/opinionated tone, adaptive thinking replaces fixed budget_tokens
Teams relying on long-context multi-needle retrieval should retain Opus 4.6 as a fallback. For coding, vision, and agentic workflows, 4.7 is the clear upgrade.
コミュニティ評価
## Community Reception
**Developer sentiment is strong but nuanced.** Anthropic's launch announcement includes testimonials from 28+ companies, and the broader community echoes the coding gains while flagging real costs and regressions.
**Positive reactions:**
- Cursor reported 70% on CursorBench vs 58% for Opus 4.6 — 'a meaningful jump in capabilities'
- Replit called it 'an easy upgrade decision,' citing 'same quality at lower cost' and describing it as pushing back during technical discussions 'like a better coworker'
- Notion reported '+14% over Opus 4.6 at fewer tokens and a third of the tool errors' — calling it 'the reliability jump that makes Notion Agent feel like a true teammate'
- Vercel noted new behavior: 'It even does proofs on systems code before starting work, which is new behavior we haven't seen from earlier Claude models'
- XBOW reported a visual-acuity benchmark jump from 54.5% to 98.5%: 'Our single biggest Opus pain point effectively disappeared'
- Rakuten measured 3× more production tasks resolved on their SWE-bench variant
**Concerns and criticisms:**
- Multiple reviewers (Awesome Agents, MindStudio, AI Fire) flagged the 10–35% token cost inflation from the new tokenizer as a real budget concern for high-volume API users
- The MRCR v2 long-context regression drew significant attention; Anthropic acknowledged it but argued MRCR doesn't reflect real-world usage patterns
- Prose quality reportedly declined — 'more mechanical, reaches for bullet points, paragraphs feel chopped' (Awesome Agents)
- Claude Code has been 'over-eager to flag benign code, including static HTML and CSS, as potential malware' — a tuning issue expected to be patched
- Some Pro/Max subscribers report hitting plan limits much faster due to higher token usage at default xhigh effort
- Community noted the removal of visible reasoning output in the web interface, breaking pipelines that depended on watching Claude's thinking process
**Adoption patterns:** Engineering teams running production code agents and autonomous workflows are adopting quickly. Research-heavy and writing-focused users are more cautious, with many keeping Opus 4.6 available for retrieval and long-context tasks. The consensus: a clear upgrade for coding/vision, a mixed bag for everything else.
ユースケース
## Specific Use Cases
### 1. Agentic Software Engineering (Code Agents, CI/CD Automation)
**When to choose Opus 4.7:** This is the model's sweet spot. It handles multi-file edits, long-running debugging sessions, and autonomous code review with significantly fewer tool errors than any competitor. Cursor's 70% CursorBench score and Rakuten's 3× production task resolution rate confirm this. Use the new xhigh effort level and task budgets to bound runaway sessions. Ideal for teams running Devin, Factory Droids, Replit Agent, or custom CI/CD agents that span tools, codebases, and debugging context.
### 2. Computer-Use Agents and Visual Document Analysis
**When to choose Opus 4.7:** The 3.3× resolution increase (3.75MP) and +13 point visual reasoning gains make it the clear leader for screenshot-heavy workflows. XBOW's autonomous penetration testing now works at 98.5% visual acuity. Use cases include: automated UI testing from dense screenshots, financial dashboard analysis, complex diagram extraction (architecture schematics, patent drawings, chemical structures), and scanned PDF processing. For OCR-heavy pipelines with fine print or dense layouts, Opus 4.7 is meaningfully better than any competitor.
### 3. Enterprise Document Analysis and Financial Modeling
**When to choose Opus 4.7:** OfficeQA Pro jumped from 57.1% to 80.6%. Harvey's BigLaw Bench score of 90.9% and the state-of-the-art 64.4% Finance Agent score confirm strength across legal and financial document reasoning. Choose this model for: multi-document due diligence, earnings report analysis, contract review (it correctly distinguishes assignment provisions from change-of-control provisions), and structured financial modeling. Databricks reports 21% fewer errors on enterprise document analysis versus Opus 4.6.
### 4. Multi-Tool Orchestration and Complex Agent Pipelines
**When to choose Opus 4.7:** At 77.3% on MCP-Atlas — 9+ points ahead of GPT-5.4 — Opus 4.7 is the strongest model for workflows that coordinate across multiple tools, APIs, and data sources. Choose it for: RAG pipelines with tool-calling, retrieval-augmented slide creation, orchestrator agents that manage sub-agents, and any workflow where the model needs to plan across many steps while maintaining consistency. Notion and Hebbia both report double-digit improvements in tool-call accuracy and planning.
**When NOT to choose Opus 4.7:**
- Web research and browsing agents (BrowseComp 79.3% trails GPT-5.4 at 89.3%)
- Long-context multi-needle retrieval over 256K tokens (use Opus 4.6 or a specialized model)
- Competition-grade math reasoning (around 70% on USAMO 2026, trailing GPT-5.4)
- Budget-sensitive high-volume API workloads (tokenizer inflates costs 10–35%)
最新ニュース
## Recent Developments (as of April 2026)
- **April 16, 2026:** Claude Opus 4.7 launched globally on Claude.ai, Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Pricing unchanged at $5/$25 per 1M input/output tokens.
- **April 16, 2026:** Anthropic published the 232-page Claude Opus 4.7 System Card covering safety evaluations, alignment testing, welfare assessments, and RSP checks. The model launched under AI Safety Level 3.
- **April 16, 2026:** New features launched alongside: xhigh effort level, task budgets (public beta), /ultrareview command in Claude Code, auto mode extended to Max users, adaptive thinking replaces fixed budget_tokens parameter.
- **April 16, 2026:** Breaking API changes: extended thinking budgets removed (setting budget_tokens returns 400 error); temperature, top_p, and top_k cannot be set to non-default values; thinking content omitted from responses by default (opt in with display: 'summarized').
- **April 16, 2026:** Cyber Verification Program launched for security professionals to access Opus 4.7 for legitimate vulnerability research, penetration testing, and red-teaming without triggering default refusal behavior.
- **April 18, 2026:** Partner benchmark results published by Cursor (70% CursorBench) and Rakuten (3× tasks resolved on SWE-bench variant).
- **April 2026:** Anthropic acknowledged tokenizer change increases token counts by 1.0–1.35× and published a migration guide with tuning advice for teams upgrading from Opus 4.6.
- **May 4, 2026:** Anthropic updated the Document reasoning graph in the launch announcement to reflect corrected OfficeQA Pro scores for Opus 4.7.
- **Context:** Claude Mythos Preview remains Anthropic's most capable model but is not broadly released. Opus 4.7 is the first model to ship with Project Glasswing cybersecurity safeguards, tested before a potential broader Mythos release.