Anthropic unveils civilian Mythos amid safety debate

What’s Claude Opus 4.7?

Meet Claude Opus 4.7, Anthropic’s newest generally available AI model, released in April 2026. It brings stronger coding performance, higher-resolution image understanding, and improved verification behavior for complex professional tasks.

Anthropic says Opus 4.7 is not as cyber-capable as Claude Mythos Preview, its more restricted model used in defensive cybersecurity work. The company says it experimented with reducing Opus 4.7’s cyber capabilities and added safeguards to block high-risk cybersecurity requests.

The model too risky for you

The more restricted model is called Claude Mythos Preview, and Anthropic announced it as a gated research preview for Project Glasswing. Anthropic describes Mythos Preview as its most capable model for coding and agentic tasks, with unusually strong vulnerability-discovery abilities.

Access is limited to Project Glasswing participants, including major technology and security partners such as AWS, Google, Microsoft, Apple, NVIDIA, JPMorgan Chase, CrowdStrike, Cisco, Palo Alto Networks, the Linux Foundation, and additional critical software organizations.

Opus 4.7 is the broadly available model, while Anthropic says lessons from its new safeguards will inform any future wider deployment of Mythos-class systems.

Good at work, bad at hacking

So what can Opus 4.7 actually do for you? It handles tough coding jobs that used to need your constant supervision. You can hand it a project and trust it to finish without you looking over its shoulder every minute. Its vision also got a major boost, reading messy diagrams and screenshots like a human would.

But here is the catch. The company built automatic guards that block any request that looks like hacking or cyberattacks. If you ask it to break into a system, it simply says no. This is the first live test of their “Project Glasswing” safety system, and the guards are very, very touchy.

The quiet price increase

Here is something the company didn’t advertise loudly. Opus 4.7 uses a new system that counts your words differently. For the same question you typed before, you might now use up to 35% more tokens. Tokens are how they measure and charge you.

The dollar price per token stayed the same on paper. But because you use more tokens for the same work, your bill quietly goes up. Developers call this a stealth price hike. Imagine paying the same for a gallon of milk but getting only two-thirds of a gallon.

The overly cautious cop

Users quickly reported that Opus 4.7’s safety filters were blocking some legitimate work. The Register reported more than 30 complaints in April about alleged false positives involving security, general development, and science-related tasks.

Examples included Russian-language prompts, computational structural biology work, and a simple crypto exercise in an LSU cybersecurity lab. The complaints suggest the new safeguards may be useful for blocking risky requests, but they can also interfere with harmless research and development work.

The 47% problem that shocked everyone

Some outside security experts have raised concerns about Claude’s coding reliability after the Opus 4.7 rollout. Forbes reported criticism from TrustedSec CEO Dave Kennedy, while Anthropic separately acknowledged product-level Claude Code issues that hurt coding quality and said the fixes were in place by April 20.

Fortune reported findings from Veracode that Opus 4.7 introduced vulnerabilities in 52% of tested coding tasks, while OpenAI models were around 30% in the same analysis. Those findings do not mean Opus 4.7 is unusable for every coding task, but they do mean generated code still needs careful review and security testing.

Little-known fact: Veracode’s testing over the last year shows OpenAI’s models perform notably better at around 30% vulnerable code, while Opus 4.7 hit 52%. 

You must prove you are nice

Want to use the AI for serious security research? You now need to apply for permission. Anthropic created a Cyber Verification Program, where professionals must prove their identity and good intentions before getting full access to certain features. It is a background check for AI.

But here is the problem. Even if you get approved, the exemption only works in the chat version, not through the coding tools professionals actually use. So you could be officially blessed as a good guy, and your work software would still refuse to help you.

They took away the controls

Previous AI models let you adjust dials like temperature and creativity to fine-tune how they worked. Opus 4.7 removed many of those controls. Try to use the old settings now, and you just get a 400 error message. The company decided they know better than you do.

The AI’s thinking process is also hidden by default. You ask a question, and the answer just appears. You cannot see whether it really thought hard or just guessed. It is like buying a car where the manufacturer welded the hood shut so you cannot check the engine.

Not everyone is complaining

Not everyone is unhappy with the new model. Replit President Michele Catasta said Opus 4.7 felt like a better coworker for Replit’s use cases, while Cursor CEO Michael Truell said it rose from 58% to 70% on CursorBench.

If you do long, complex tasks and are willing to re-train your workflow, the new model might actually be better. But if you expected a simple upgrade that just works out of the box, you will probably be disappointed. Your experience depends entirely on how you use it.

Little-known fact: On the SWE-bench Verified test, which measures a model’s ability to solve real-world GitHub issues, Opus 4.7 achieved 87.6%, a nearly 7-point jump from Opus 4.6’s 80.8%.

The two-tier internet is here

The launch shows a growing split between broadly available AI models and restricted frontier systems for sensitive domains. Anthropic’s Mythos Preview remains gated through Project Glasswing, while Opus 4.7 is the generally available model with additional cybersecurity safeguards.

That access model raises real questions about who gets the most powerful AI capabilities and under what conditions. Anthropic says the goal is to support defensive cybersecurity and eventually enable safer deployment of Mythos-class systems at scale.

What the boss says

Anthropic said in an official April 23 postmortem that it “never intentionally degrade[s]” its models and traced recent Claude Code quality complaints to three product-level issues. The company said those issues affected Claude Code, the Claude Agent SDK, and Claude Cowork, while the API and inference layer were unaffected.

Anthropic said the issues were fixed as of April 20 and included a default-effort change, a caching bug, and a system-prompt change that hurt coding quality. Even so, user frustration around model behavior, token usage, and safety filters has continued to shape the response to Opus 4.7.

The new thinking mode

Opus 4.7 introduced a new xhigh effort level between high and max, giving users another way to balance reasoning depth, latency, and token use. Anthropic says Claude Code now defaults to xhigh for Opus 4.7 across all plans.

The change can improve performance on harder coding and agentic tasks, but it can also increase token usage in some workflows. Developers should check effort settings, task budgets, and token counts before assuming Opus 4.7 will cost the same per job as Opus 4.6.

Want to see where concerns about AI behavior are heading next? Take a look at the spotlight on Mythos and rising AI crime risks.

Should you still use it?

For everyday writing, summarization, brainstorming, and light research, Opus 4.7 may work well for many users. The strongest caution is for people who rely on it for coding, cybersecurity, or other high-stakes professional workflows.

Before switching, compare it against the older model on your own tasks, review token usage, and check outputs carefully. Opus 4.7 is powerful, but the launch has shown that stronger benchmarks, new safeguards, and real-world reliability do not always move in lockstep.

Want a bigger-picture take on where Anthropic is headed? Check out why some think it could come out ahead in its fight with Trump; it’s an interesting angle.

Got a strong opinion about this civilian AI or its secret big brother? Drop a comment below, and give this slideshow a thumbs up if you found it helpful.

This slideshow was made with AI assistance and human editing.

Don’t forget to follow us for more exclusive content on MSN.

Read More From This Brand:

• Microsoft faces backlash for 6000 visa push amid job cuts
• Microsoft experiments with Copilot AI that can operate across your desktop
• Microsoft faces scrutiny over alleged Chinese engineers handling DOD systems