Anthropic released Claude Opus 4.7 on Thursday, April 16, 2026 — its most capable generally available AI model and the direct successor to Claude Opus 4.6, which launched in February. The release addresses months of user complaints that Opus 4.6 had quietly degraded, delivers measurable performance gains on coding and agentic tasks, and introduces the first real-world deployment of cyber safeguards designed to eventually enable a broader release of Mythos-class capabilities.
For developers who have been waiting, or who migrated away from Opus 4.6 due to performance concerns: Opus 4.7 is a meaningful upgrade, not an incremental patch.
Direct Answer: What is Claude Opus 4.7 and what’s new?
Claude Opus 4.7 is Anthropic’s latest and most capable generally available AI model, released April 16, 2026. Key improvements over Opus 4.6: 13% coding benchmark lift, self-verification during planning phase, better image vision, stronger instruction following, file system memory for multi-session projects, and a new xhigh effort level for finer reasoning control. It outperforms Opus 4.6, GPT-5.4, and Gemini 3.1 Pro — but sits below the restricted Claude Mythos Preview. Pricing unchanged: $5/M input, $25/M output. Available now via claude-opus-4-7 on the API, Amazon Bedrock, Vertex AI, Microsoft Foundry, and Claude Pro/Max/Team/Enterprise.
What’s Actually New in Opus 4.7
1. Coding: 13% Benchmark Lift and Self-Verification
The headline number from Hex’s independent 93-task coding benchmark: Opus 4.7 lifted resolution by 13% over Opus 4.6, including four tasks that neither Opus 4.6 nor Sonnet 4.6 could solve at all.
Anthropic’s description of the qualitative change: “Users report being able to hand off their hardest coding work — the kind that previously needed close supervision — to Opus 4.7 with confidence.”
The most significant new behaviour is self-verification during the planning phase. Before executing a plan, Opus 4.7 now checks its own logical reasoning for faults and corrects them before proceeding. This is described by Anthropic as the model catching “its own logical faults during the planning phase and accelerates execution, far beyond previous Claude models.”
In practice this means fewer dead-end executions on complex multi-step tasks. Previous models would begin executing a subtly flawed plan and produce output that needed significant correction. Opus 4.7 catches many of these faults before they become downstream errors.
SWE-bench and agentic benchmarks: Anthropic reports Opus 4.7 beats Opus 4.6, GPT-5.4, and Gemini 3.1 Pro across industry benchmarks for:
- Agentic coding (SWE-bench and variants)
- Multidisciplinary reasoning
- Scaled tool use
- Agentic computer use
GitHub Copilot adoption: GitHub confirmed Opus 4.7 is rolling out on GitHub Copilot Pro+, where it will replace Opus 4.5 and 4.6 over the coming weeks. GitHub’s internal testing shows “stronger multi-step task performance and more reliable agentic execution” and “meaningful improvement in long-horizon reasoning and complex, tool-dependent workflows.”
2. The xhigh Effort Level
This is the most practically significant new control for power users and developers.
Opus 4.7 introduces a new effort level: xhigh (“extra high”) — positioned between the existing high and max levels. It gives users and developers finer control over the tradeoff between reasoning depth and latency on hard problems.
Anthropic’s direct recommendation: “When testing Opus 4.7 for coding and agentic use cases, we recommend starting with high or xhigh effort.”
The practical implications:
higheffort: best for most tasks — good balance of quality and speedxhigheffort: for complex, multi-step problems where additional reasoning time produces meaningfully better resultsmaxeffort: maximum reasoning, highest latency and token cost, for the hardest tasks
Also new: task budgets — a system that gives developers more control over how Claude allocates reasoning time across different parts of longer tasks. This is designed for agentic workflows where some subtasks need deep reasoning and others do not.
3. Better Vision
Opus 4.7 has substantially better image understanding — it can analyse images at higher resolution than Opus 4.6. Anthropic describes this as being able to “see images in greater resolution.”
For developers working with multimodal workflows (code that involves screenshots, diagrams, or visual assets), this is a direct improvement. For users asking Claude to analyse charts, documents, or interface screenshots, image comprehension quality improves meaningfully.
4. File System Memory Across Sessions
Opus 4.7 is significantly better at using file system-based memory to carry context across long, multi-session projects.
From Anthropic’s blog: “It remembers important notes across long, multi-session work, and uses them to move on to new tasks that, as a result, need less up-front context.”
This directly addresses one of the most common complaints about previous Claude models in enterprise workflows: the need to re-establish project context at the start of every new conversation. Opus 4.7 can persist relevant context in structured memory files, access them on subsequent sessions, and carry state across what are effectively multi-day projects without requiring users to re-brief the model each time.
5. Instruction Following and Professional Output Quality
Anthropic describes “precise attention to instructions” and improved “taste and creativity when completing professional tasks, producing higher-quality interfaces, slides, and docs.”
Box’s Head of AI Yashodha Bhavnani confirmed in early testing: “Claude Opus 4.7 demonstrates significant efficiency gains while preserving the performance of Claude Opus 4.6.” The Box evaluation found improved instruction following specifically in enterprise document workflows.
Hex’s evaluation noted that Opus 4.7 “correctly reports when data is missing instead of providing plausible-but-incorrect fallbacks, and it resists dissonant-data traps that even Opus 4.6 falls for” — a direct improvement in the model’s honesty under uncertainty.
6. Tokeniser Update — Important for Developers
This is a breaking change developers need to plan for.
Opus 4.7 uses an updated tokeniser that changes how text maps to tokens. The tradeoff: the same input can map to more tokens — roughly 1.0–1.35× depending on content type. Additionally, Opus 4.7 thinks more at higher effort levels, particularly on later turns in agentic settings, producing more output tokens.
Practical impact: If you have cost budgets or token limits set based on Opus 4.6 usage patterns, you will need to adjust them upward by approximately 15–35%. The efficiency improvement (low-effort Opus 4.7 roughly equals medium-effort Opus 4.6) may offset this in many workflows, but it requires testing.
Benchmarks: Where Opus 4.7 Stands
Anthropic published a benchmark comparison chart with the Opus 4.7 release. The confirmed positioning:
| Model | Agentic Coding | Reasoning | Computer Use | Available |
|---|---|---|---|---|
| Claude Mythos Preview | Best | Best | Best | ❌ Project Glasswing only |
| Claude Opus 4.7 | ✅ Frontier | ✅ Frontier | ✅ Frontier | ✅ Generally available |
| Claude Opus 4.6 | Below 4.7 | Below 4.7 | Below 4.7 | Still available |
| GPT-5.4 | Below Opus 4.7 | Below Opus 4.7 | — | Generally available |
| Gemini 3.1 Pro | Below Opus 4.7 | Below Opus 4.7 | — | Generally available |
The honest context on Mythos: Anthropic is transparent that Opus 4.7 is “less broadly capable” than Mythos Preview. Mythos has AISI-verified capabilities to complete full 32-step enterprise network attacks and discover previously unknown vulnerabilities. Opus 4.7 does not. Mythos will not be publicly available; Opus 4.7 is. For practical purposes, Opus 4.7 is the best AI model a developer or business can actually deploy.
Pricing and Availability
Pricing: Unchanged from Opus 4.6.
- Input: $5 per million tokens
- Output: $25 per million tokens
- With prompt caching: up to 90% cost savings
- With batch processing: 50% savings
- US-only inference: 1.1× pricing on input and output tokens
Model string: claude-opus-4-7
Available on:
- Claude.ai (Pro, Max, Team, Enterprise plans)
- Claude API (direct)
- Amazon Bedrock
- Google Cloud Vertex AI
- Microsoft Foundry (Azure)
- GitHub Copilot Pro+ (rolling out over the coming weeks, replacing Opus 4.5 and 4.6)
Promotional note on GitHub Copilot: GitHub is launching Opus 4.7 with a 7.5× premium request multiplier as part of promotional pricing until April 30, 2026.
The Cyber Safeguards Story
Opus 4.7 is not just a performance upgrade. It is also Anthropic’s first real-world deployment of a new cyber safety framework designed to eventually enable a broader release of Mythos-class capabilities.
From Anthropic’s official announcement: “Opus 4.7 is the first such model: its cyber capabilities are not as advanced as those of Mythos Preview (indeed, during its training we experimented with efforts to differentially reduce these capabilities). We are releasing Opus 4.7 with safeguards that automatically detect and block requests that indicate prohibited or high-risk cybersecurity uses.”
This is technically significant. Anthropic deliberately trained Opus 4.7 to have reduced cyber exploit capabilities relative to what its base architecture would otherwise produce. The “differential capability reduction” during training — rather than just adding post-training filters — represents a new approach to AI safety: shaping the model’s learned capabilities at the architectural level rather than constraining them after the fact.
The practical deployment of automatic detection-and-blocking safeguards on Opus 4.7 is a testing ground. If these safeguards work reliably at scale on a model with deliberately reduced (but still significant) cyber capabilities, Anthropic can apply the same approach to eventually enable broader access to Mythos-class capability levels.
The Cyber Verification Program: Security professionals who want to use Opus 4.7 for legitimate cybersecurity purposes — vulnerability research, penetration testing, red-teaming — can apply to a new Cyber Verification Program that grants additional access levels not available to general API users.
The “Nerfed 4.6” Controversy: Addressed
Before unpacking the upgrade, it is worth addressing the context.
For weeks before the Opus 4.7 release, vocal developer complaints circulated on GitHub, Hacker News, and X that Opus 4.6 had quietly degraded. An AMD senior director wrote in a widely shared GitHub post: “Claude has regressed to the point it cannot be trusted to perform complex engineering.” The speculation was that Anthropic had deliberately scaled back Opus 4.6 — “nerfed” it — either to reduce compute costs or to redirect capacity toward Mythos and Opus 4.7 development.
Anthropic denied deliberately redirecting compute. But the complaints were real, documented, and came from credible enterprise users who had benchmarked the model systematically over time.
The Axios report on the Opus 4.7 launch directly acknowledged this context: “The release arrives amid weeks of user complaints that Opus 4.6 had quietly gotten worse.”
Opus 4.7’s arrival does not retroactively explain what happened to 4.6 — Anthropic has not published a detailed explanation. But it does provide a resolution: the model developers were frustrated with is being replaced by one that demonstrably performs better on the tasks where 4.6 fell short.
How to Use Claude Opus 4.7 in Claude Code
For Claude Code users specifically, several changes are relevant:
1. Update your model string. Change from claude-opus-4-6 to claude-opus-4-7 in your project configuration.
2. Use xhigh for hard tasks. For your most complex engineering problems — the ones you previously found needed heavy supervision — start with xhigh effort. Anthropic recommends this as the sweet spot for difficult agentic coding.
3. Expect more tokens. The tokeniser update means the same prompts will cost more. Run a small benchmark of your typical workloads before switching your entire production setup.
4. Enable file system memory. For multi-session projects, configure Claude to write and read memory files. The improved memory architecture in Opus 4.7 makes this a much more reliable workflow than with previous models.
5. Use the new /ultrareview command. Claude Code now ships a /ultrareview command that “runs a dedicated review session that reads through your changes and flags what a careful reviewer would catch.” This is separate from Opus 4.7 itself but was released alongside it.
6. Task budgets for long workflows. For agentic tasks that span many steps, experiment with task budgets to give the model appropriate reasoning allocation across different phases of the work.
Claude Opus 4.7 and the Sovereign AI Stack
From Vucense’s perspective, the most important aspect of Opus 4.7 is what it means for the sovereign local AI picture.
What Opus 4.7 is not: It is not a local model. It runs on Anthropic’s infrastructure (or Amazon, Google, Microsoft). Your prompts and outputs pass through external servers. For sensitive workloads, this remains a data sovereignty concern regardless of how capable the model becomes.
What it means for the hybrid stack: The improvements in multi-session memory, instruction following, and self-verification make Opus 4.7 a significantly more capable orchestrator in a hybrid architecture — where Opus 4.7 handles complex reasoning and planning, while local models (Llama 4, Gemma 4, or TurboQuant-compressed models when they arrive) handle sensitive data processing that cannot leave your infrastructure.
The API access layer: Unlike models that require proprietary platforms, Opus 4.7 is accessible via standard API with pricing transparency and multi-cloud availability. This means you can route through Amazon Bedrock (US-hosted infrastructure, data residency options) or Google Vertex AI, applying your own data governance policies to API calls rather than relying solely on Anthropic’s defaults.
sovereignScore: 80 — Opus 4.7 scores well for API transparency, pricing clarity, US data residency options, and multi-cloud deployment. It scores lower for the fundamental limitation of all cloud AI: your data leaves your device. For teams handling regulated data, client confidential information, or security-sensitive workloads, the Cyber Verification Program’s access controls and the US-only inference option are relevant tools to explore.
FAQ
How do I access Claude Opus 4.7?
Use claude-opus-4-7 as your model string via the Anthropic API. It is also available on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. For Claude.ai users, Opus 4.7 is available on Pro, Max, Team, and Enterprise plans starting today.
What is the price of Claude Opus 4.7? Unchanged from Opus 4.6: $5 per million input tokens, $25 per million output tokens. Up to 90% savings with prompt caching, 50% with batch processing.
Is Claude Opus 4.7 better than GPT-5.4? Anthropic’s benchmark data shows Opus 4.7 outperforms GPT-5.4 on agentic coding, multidisciplinary reasoning, and scaled tool use. Independent evaluations (Hex’s 93-task coding benchmark showing 13% improvement over Opus 4.6) support this positioning. As always with benchmarks, performance varies by task type — your specific use case is the real test.
What is the xhigh effort level?
A new reasoning intensity option between high and max. Use it for complex, multi-step problems where you want more reasoning depth without going to maximum latency. Anthropic recommends high or xhigh as starting points for coding and agentic workflows.
What happened to the tokeniser? Opus 4.7 uses an updated tokeniser. The same input text maps to approximately 1.0–1.35× more tokens than with Opus 4.6. Adjust cost budgets accordingly before switching production workloads.
Is Claude Opus 4.7 the same as Claude Mythos? No. Mythos is Anthropic’s most powerful model but is not publicly available — it is restricted to Project Glasswing partners (Apple, Google, Microsoft, JPMorgan, etc.) due to cybersecurity concerns. Opus 4.7 is the best model that anyone can actually use. Anthropic explicitly states Opus 4.7 is “less broadly capable” than Mythos.
What is the Cyber Verification Program? A new Anthropic programme for security professionals (vulnerability researchers, pen testers, red teams) who want to use Opus 4.7 for legitimate cybersecurity work. Standard API access includes safeguards that block high-risk cybersecurity requests; verified professionals can apply for expanded access.
Why did Claude Opus 4.6 seem to get worse before 4.7? Anthropic denied deliberately redirecting compute. Enterprise users documented real performance degradation in the weeks before the 4.7 release. Anthropic has not published a detailed explanation. Opus 4.7 replaces 4.6 and demonstrably performs better on the tasks where complaints centred.
Related Articles
- Claude Mythos: The AI That Scared the Fed, the Bank of England, and Every CISO on the Planet
- Claude Code + TurboQuant: Run 70B Models Locally Without Cloud
- TurboQuant + Ollama & llama.cpp: 2026 Setup Guide
- Anthropic Overtakes OpenAI: $30 Billion ARR, IPO October 2026
- Claude Managed Agents: Anthropic Launches Enterprise AI Agent Infrastructure
- Cursor vs GitHub Copilot vs Claude Code 2026: Full Audit
Sources & Further Reading
- MIT Technology Review — AI Section — In-depth coverage of AI research and industry trends
- arXiv AI Papers — Pre-print research papers on AI and machine learning
- EFF on AI — Civil liberties perspective on AI policy
