Anthropic Releases Claude Opus 4.7 and Previews 10-Trillion Parameter Mythos Model

On April 16, 2026, Anthropic made two major announcements that are reshaping the AI landscape: the general availability of Claude Opus 4.7, its new flagship model, and a limited preview of Claude Mythos 5 — a staggering 10-trillion parameter model that Anthropic considers too powerful for public release. Together, these releases reveal a company walking a careful line between advancing AI capabilities and managing the profound risks that come with them.

Claude Opus 4.7: What’s New

The Anthropic Claude Opus 4.7 release represents a significant iterative upgrade to Anthropic’s publicly available model series, designed to handle complex, long-running tasks for enterprise users and developers. Following the company’s predictable two-month upgrade cadence, Opus 4.7 focuses on three core areas: advanced software engineering, enhanced vision, and enterprise-grade reliability.

Advanced Software Engineering

Opus 4.7 exhibits notable improvements in advanced software engineering, allowing developers to confidently delegate complex coding jobs that previously required close human supervision. The model handles long-running tasks with rigor, demonstrates precise instruction following, and can verify its own outputs. A new “xhigh” effort level gives users finer control over the trade-off between reasoning depth and latency for challenging problems.

Enhanced Vision Capabilities

The model’s vision capabilities have been substantially upgraded. Opus 4.7 can process images with a resolution of up to 2,576 pixels on the long edge — approximately 3.75 megapixels, more than triple the resolution of its predecessors. This enables more detailed visual analysis, such as parsing dense screenshots for computer-use agents and extracting structured data from complex technical diagrams.

Enterprise Safety and Reliability

Anthropic has designed Opus 4.7 to address common enterprise challenges like model drift and hallucinations. The model features “loop resistance” to avoid repetitive cycles and is programmed to push back against generating false information. It also shows improvements in honesty and resistance to prompt-injection attacks. Anthropic has deliberately reduced the model’s inherent cyber capabilities and implemented safeguards to block high-risk cybersecurity uses, with legitimate security professionals encouraged to apply for a Cyber Verification Program.

📖 Related: using Claude for AI-powered code review

Benchmark Performance: How Does It Stack Up?

Claude Opus 4.7 demonstrates a competitive edge in specific domains, particularly agentic coding and tool use, when compared to contemporary models like GPT-5.4 and Gemini 3.1 Pro.

Where Opus 4.7 Leads

• SWE-bench Pro (full engineering pipeline): 64.3% — ahead of GPT-5.4 (57.7%) and Gemini 3.1 Pro (54.2%)
• MCP-Atlas (scaled tool use): 77.3% — highest among frontier models
• Finance Agent v1.1 (multi-step financial analysis): 64.4%
• CharXiv Reasoning (visual reasoning with tools): 91.0% — up from 84.7% in its predecessor
• Humanity’s Last Exam (multimodal, without tools): 46.9%

Where Competitors Lead

• GPQA Diamond (graduate-level scientific reasoning): All three models are nearly tied (~94%), suggesting benchmark saturation
• MMMLU (multilingual Q&A): Gemini 3.1 Pro leads at 92.6%
• BrowseComp (agentic web research): Opus 4.7 scores 79.3%, trailing GPT-5.4 Pro (89.3%) and Gemini 3.1 Pro (85.9%)

The benchmark picture reveals a highly specialized competitive landscape: no single model dominates across all tasks, and the choice of model increasingly depends on specific use case requirements.

Pricing and Availability

Claude Opus 4.7 is broadly available across all of Anthropic’s products — Pro, Max, Team, and Enterprise plans — as well as its API and through cloud providers including Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Foundry. The pricing remains unchanged from its predecessor at $5 per million input tokens and $25 per million output tokens.

However, there’s an important caveat: due to a new tokenizer, users may see their token counts for the same text increase by 1.0x to 1.35x, effectively increasing costs for some workloads. Businesses should factor this into their cost projections when migrating from Opus 4.6.

📖 Related: OpenAI’s GPT-5.4 — Anthropic’s main competitor

Claude Mythos 5: A Glimpse at the Future

While Opus 4.7 is a capable upgrade, Anthropic has made it clear that it is “less broadly capable” than its preview model, Claude Mythos 5. Internally codenamed “Capybara,” Mythos represents a step-change in AI capabilities that has captured the attention of the entire industry.

Scale and Architecture

Mythos 5 is estimated to be a 10-trillion parameter model — a significant leap from the estimated 1.8 trillion parameters of models like GPT-4. It utilizes a sophisticated Mixture-of-Experts (MoE) architecture, where only a fraction of parameters (estimated 800 billion to 1.2 trillion) are active during any single task. This allows it to retain the vast knowledge of a 10T model while keeping computational costs comparable to a much smaller dense model. It also features a rumored context window of up to 4 million tokens, enabling deep synthesis of vast amounts of information.

Scientific Reasoning

Beyond cybersecurity, Mythos excels in academic reasoning, achieving a 97.2% on the USAMO math olympiad benchmark and demonstrating an ability to identify methodological weaknesses in scientific research papers — a capability that could transform peer review and research validation.

Unprecedented Cybersecurity Capabilities

The most striking — and concerning — capabilities of Mythos 5 lie in cybersecurity. Early access testing has shown the model can autonomously perform full attack chain analysis and identify potential zero-day vulnerabilities in source code. In one documented instance, it discovered a 27-year-old bug in OpenBSD that had previously eluded human experts. It can autonomously write exploits for identified vulnerabilities, demonstrating a 90x improvement over Opus 4.6 on one benchmark.

Anthropic has noted these skills were an emergent property of the model’s general improvements in reasoning and code — not a direct training objective. This emergent capability is precisely why the company has chosen not to make Mythos publicly available.

📖 Related: Google’s Gemini 3.1 — another frontier model competitor

Project Glasswing: Controlled Access Only

Anthropic’s concerns over Mythos 5’s potent offensive cyber capabilities have led to a highly restricted rollout under Project Glasswing. Limited access is provided to a select group of organizations — including Apple, Microsoft, Google, and CrowdStrike — strictly for defensive cybersecurity work. There is no planned general release date.

Mythos introduces a new, highly expensive “Capybara” model tier, with early access pricing estimated to be substantially higher than the Opus tier, targeting well-capitalized enterprise and government clients. This two-tier strategy — a capable public model and a restricted frontier model — may become a template for how AI companies manage the deployment of increasingly powerful systems.

Market Reaction and Strategic Implications

Expert reaction has focused on Anthropic’s dual-pronged strategy. Opus 4.7 is seen as a solid, competitive upgrade that keeps pace with rivals like OpenAI and Google. However, the preview of Mythos 5 has captured the most attention, highlighting both the immense potential of next-generation AI and the profound safety challenges that come with it.

Anthropic’s decision to restrict access to Mythos is widely viewed as a responsible, safety-first approach to deploying such a powerful technology. It also serves a strategic purpose: by previewing Mythos, Anthropic signals to the market that it is operating at the frontier of AI research, even if that frontier is not yet accessible to the public.

Conclusion

The Anthropic Claude Opus 4.7 release and the preview of Claude Mythos 5 together tell a compelling story about the state of AI in 2026. Opus 4.7 is a powerful, enterprise-ready model that leads in coding and agentic tasks. Mythos 5 is a glimpse of a future where AI capabilities may outpace our ability to safely deploy them.

For businesses and developers, Opus 4.7 offers immediate, practical value. For the broader AI community, Mythos 5 raises urgent questions about how we govern, access, and ultimately coexist with AI systems of extraordinary power. Anthropic’s approach — advance carefully, restrict responsibly — may well define the industry’s standard for years to come.