Anthropic shipped Claude Opus 4.7 on April 16, 2026, with measurable improvements in coding capability and vision resolution, though the company is deliberately withholding a more powerful model called Mythos Preview due to cybersecurity concerns. The new version shows a jump from 58% to over 70% on CursorBench, a real-world coding benchmark, and handles up to 2576 pixels on the long edge of images, roughly 3.75 megapixels. Pricing remains flat at $5 per million input tokens and $25 per million output tokens, though a new tokenizer can increase token counts by up to 35% on identical workloads .

What Makes Opus 4.7 Different From Its Predecessor?

The jump from Opus 4.6 to 4.7 raised eyebrows because point-release updates typically don't deliver such dramatic performance gains. The improvements are grounded in production workloads rather than internal benchmarks alone. Cursor's CEO Michael Truell confirmed the 12-point CursorBench improvement, while Rakuten's general manager Yusuke Kaji reported that the model resolved three times more production bugs compared to Opus 4.6 on Rakuten-SWE-Bench, a benchmark designed around real-world software engineering tasks .

Vision capabilities expanded significantly. The model can now process images up to 3.75 megapixels, enabling it to read small text in diagrams, user interface screenshots, and slides. This represents a threefold improvement over prior versions. Anthropic also introduced three new features to help developers manage performance and cost :

• xhigh effort level: A new reasoning tier between high and max that lets developers balance thinking depth against response speed when max is too slow but high isn't sufficient.
• Task budgets (public beta): Allows developers to set token spending limits during long agent runs, prioritizing expensive reasoning steps while economizing on routine operations.
• Improved tokenizer: Better text handling, though the same input can now produce 1.0 to 1.35 times more tokens, potentially increasing invoices despite unchanged per-token pricing.

Why Is Anthropic Holding Back Mythos Preview?

The most intriguing aspect of this release is what Anthropic is not shipping publicly. The company explicitly stated that Opus 4.7 had its cybersecurity capabilities deliberately reduced. This connects to Project Glasswing, a coalition announced last week that includes Amazon Web Services (AWS), Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. The coalition focuses on securing critical software for the artificial intelligence era .

Mythos Preview, a more powerful closed-access model, runs against public and proprietary codebases through Project Glasswing. Anthropic claims Mythos has already uncovered thousands of high-severity vulnerabilities across every major operating system and browser. That capability becomes a cyber-weapon if released openly, so Mythos remains restricted. Opus 4.7 is effectively the publicly releasable version with the risky cyber capabilities stripped out. A separate "Cyber Verification Program" gives credentialed security researchers access to more powerful capabilities through a vetted channel .

How to Decide Whether to Upgrade to Opus 4.7

• Complex code generation and heavy refactoring: Upgrade to Opus 4.7. The CursorBench and Rakuten-SWE-Bench gains appear grounded in real production workloads, and tool-error reduction is measurable.
• Long-running agent tasks: Upgrade to Opus 4.7. Task budgets and fewer tool errors make this version more reliable for extended operations that require multiple steps.
• Image inputs with diagrams, user interfaces, or slides: Upgrade to Opus 4.7. The 3.75-megapixel resolution unlocks the ability to read small text that prior versions missed.
• Everyday writing, summaries, and translation: Hold off on upgrading. Sonnet 4.6 is sufficient for these tasks; Opus 4.7 is overkill and more expensive.
• High-volume batch processing or speed-sensitive applications: Use Haiku instead. Opus is heavy on both latency and cost for these use cases.
• Self-built applications on the API: Measure actual token counts on your real inputs before switching. The 1.0 to 1.35 times range is too wide to plan around blindly.

What Should You Watch Out For?

The tokenizer change deserves careful attention. Anthropic itself acknowledges that the new tokenizer can produce up to 35% more tokens for the same input. Per-token pricing didn't change, which means your bill could grow substantially on identical workloads. If you're running Opus at scale, you need to re-measure token counts against your real input distribution before switching .

GitHub Copilot users face a time-sensitive consideration. Per GitHub's changelog, Opus 4.7 on Copilot uses a 7.5 times premium request multiplier, discounted through April 30, 2026. From May onward, costs increase significantly. If you're on Copilot, burn through enough requests this month to understand your actual cost profile before the discount ends .

The "3x resolved tasks" figure comes from Rakuten's specific internal benchmark. Don't assume your workload will see a threefold lift. The real-world improvements are probably real, but the 3x multiplier is not portable across all use cases. Reasonable interpretation: the coding capability lift is backed by multiple primary sources, tool-error reduction is measurable, and 3.75-megapixel vision resolution is a genuine improvement. The headline "price unchanged" glosses over the fact that real invoices can still grow up to 35% via tokenizer changes .

Anthropic's open-and-closed split is unusual. Companies rarely announce "this isn't our best model; our best one is locked up." Anthropic explicitly states that Opus 4.7 is below Mythos Preview on general capability, but "Mythos Preview scores highest on alignment evaluations." The double message, that Mythos is more capable and safer but the deliberately capped Opus 4.7 is what should be released, represents a new kind of framing in AI governance. The Glasswing partner roster, which includes Amazon, Apple, Cisco, Google, Microsoft, and JPMorgan, is difficult to dismiss as pure marketing theater, though the setup also creates brand value .

Opus 4.7 is available across Claude, the API, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Same-day rollout across multiple platforms means environment risk is low for those doing heavy code work. For chat-style work, Sonnet 4.6 remains sufficient, and there's no urgency to upgrade. Whether the 3x jump reproduces outside Rakuten's specific benchmark will require a few months of third-party numbers to confirm .