Why Anthropic built the world's most powerful AI then chose not to release it

Anthropic just made a fascinating decision that reveals how the AI industry is grappling with its own power. They released Claude Opus 4.7 today across all Claude products and our API, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry—but kept their most capable model, Mythos Preview, locked away from public use.

This isn't just another incremental AI update. It's a glimpse into how AI companies are starting to self-regulate when their models become too powerful for comfort.

What makes Claude Opus 4.7 different?

Opus 4.7 is a notable improvement on Opus 4.6 in advanced software engineering, with particular gains on the most difficult tasks. Users report being able to hand off their hardest coding work—the kind that previously needed close supervision—to Opus 4.7 with confidence.

Claude Opus 4.7 is available through Anthropic's official platform and represents their most publicly available intelligent model for complex reasoning and coding tasks.

The improvements go beyond just coding. Opus 4.7 handles complex, long-running tasks with rigor and consistency, pays precise attention to instructions, and devises ways to verify its own outputs before reporting back. The model also has substantially better vision: it can see images in greater resolution. It's more tasteful and creative when completing professional tasks, producing higher-quality interfaces, slides, and docs.

How much better is the vision processing?

On vision, Opus 4.7 can now accept images up to 2,576 pixels on the long edge, roughly 3.75 megapixels, more than three times the resolution supported by prior Claude models. Anthropic said this expands the model's usefulness for tasks requiring fine visual detail, including reading dense screenshots and extracting data from complex diagrams.

This isn't just a technical spec bump—it's meaningful for real work. The higher resolution enables more accurate computer-use agents and better document processing capabilities.

Why did Anthropic hold back their best model?

Here's where things get interesting. Earlier today we announced Claude Mythos Preview, a new general-purpose language model. This model performs strongly across the board, but it is strikingly capable at computer security tasks.

Claude Mythos Preview is offered separately as a research preview model for defensive cybersecurity workflows as part of Project Glasswing. Access is invitation-only and there is no self-serve sign-up.

The reason for this restriction became clear during testing. We then look at Mythos Preview's ability to find and exploit zero-day (that is, undiscovered) vulnerabilities in real open source codebases. After that we discuss how Mythos Preview has proven capable of reverse-engineering exploits on closed-source software, and turning N-day (that is, known but not yet widely patched) vulnerabilities into exploits.

What can Mythos actually do?

The capabilities are genuinely concerning. Two years ago, the best available models could barely complete beginner-level cyber tasks. Now, in controlled evaluations where Mythos Preview was explicitly directed and given network access to do so, we observed that it could execute multi-stage attacks on vulnerable networks and discover and exploit vulnerabilities autonomously – tasks that would take human professionals days of work.

On expert-level tasks — which no model could complete before April 2025 — Mythos Preview succeeds 73% of the time.

Even more striking: In the controlled test, Mythos Preview autonomously surfaced thousands of "zero day" vulnerabilities – flaws unknown even to the software's own developers – across every major operating system and popular web browser.

How is Anthropic using Opus 4.7 as a testing ground?

We stated that we would keep Claude Mythos Preview's release limited and test new cyber safeguards on less capable models first. 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 a fascinating approach—using a commercially available model as a real-world testing ground for safety systems they'll eventually need for more powerful models.

Anthropic said it experimented with efforts to "differentially reduce" Claude Opus 4.7's cyber capabilities during training. The company encouraged security professionals who are interested in using the model for "legitimate cybersecurity purposes" to apply through a formal verification program.

What safeguards are actually in place?

Opus 4.7 ships with safeguards that automatically detect and block requests indicating prohibited or high-risk cybersecurity uses. Security professionals who wish to use Opus 4.7 for legitimate cybersecurity purposes, such as vulnerability research, penetration testing, and red-teaming, are invited to join Anthropic's new Cyber Verification Program.

But early reports suggest these safeguards might be overly aggressive. In April, devs have filed more than 30 reports of claimed false positives related to security, general development use, and science refusals. Issue #48442, "Persistent AUP false positives — 40+ per 4 sessions, across unrelated projects (psychology book, web app, infra, bot)," deals with Claude's refusal to process various Russian prompts. Issue #49751, "Opus 4.7 flags standard computational structural biology as Usage Policy violation, regression from 4.6," which describes computational structural biology tasks being flagged.

Should you upgrade to Opus 4.7?

The answer depends on what you're doing and how much you're willing to spend. Developers upgrading from Opus 4.6 should account for two cost-related changes. Opus 4.7 uses an updated tokenizer that can map the same input to roughly 1.0 to 1.35 times as many tokens, depending on content type. The model also produces more output tokens at higher effort levels, particularly in later turns of agentic tasks, because it engages in more reasoning.

For complex coding work, the improvements seem substantial. On the SWE-bench Pro coding benchmark, Opus 4.7 scores 64.3 percent, up from 53.4 percent for its predecessor and ahead of OpenAI's GPT-5.4 at 57.7 percent. Anthropic's own top model, Claude Mythos Preview, still leads by a wide margin at 77.8 percent.

What about instruction following?

This is where you'll need to be careful. Anthropic says Opus 4.7 follows instructions more precisely than its predecessor. The company notes that prompts written for older models may now produce unexpected results, as Opus 4.7 interprets instructions more literally than Opus 4.6, which sometimes loosely interpreted or skipped parts of them entirely.

If you're upgrading, plan to revisit your existing prompts and workflows.

What does this tell us about the future of AI?

The Opus 4.7 / Mythos split reveals something important about where AI development is heading. AI companies are starting to hold back their most capable models and limit who gets access, especially where misuse is a real concern.

Future models developed by Anthropic and other leading AI companies are being designed to function as highly autonomous AI agents, capable of independently planning, adapting and executing long, complex sequences of tasks. As well as discovering vulnerabilities, this could mean coordinating large-scale operations or managing sophisticated real-world workflows – all with minimal human guidance.

This isn't fear-mongering—it's a recognition that AI capabilities are advancing faster than our ability to safely deploy them. The fact that Anthropic built something so powerful they chose not to release it should tell us something about the trajectory we're on.

The industry is learning to self-regulate not because they want to limit innovation, but because the alternative—releasing capabilities without understanding their implications—poses genuine risks to the systems we all depend on.

For now, Opus 4.7 gives you access to significantly improved coding and vision capabilities with reasonable safety guardrails. But the shadow of Mythos hanging over this release reminds us that we're entering an era where the most powerful AI systems might not be the ones you can simply sign up and use.