On 7 April 2026, Anthropic announced Claude Mythos Preview, the most capable AI model ever built.
It outperforms most human specialists at coding, scientific reasoning, and cybersecurity research.
Anthropic chose not to release it publicly, instead assembling a coalition of competitors to deploy it safely. That decision tells you everything about what we're dealing with.
Key Takeaways
Let's start with the basics, because this isn't just a story for tech people.
Large language models (LLMs) are AI systems trained on vast amounts of text to understand and generate human language. Think of them as incredibly sophisticated pattern-matching engines that can write, code, reason, and analyse at scales humans simply can't match.
You've probably used ChatGPT, Copilot, or similar tools. Those are LLMs.
Anthropic is the company behind Claude, one of the most capable AI systems in the world. Founded by former OpenAI researchers who left to build AI more responsibly, they've consistently focused on safety alongside capability. They're not the biggest player by market cap, but they're arguably the most thoughtful.
Here's why this announcement is different from the usual AI hype cycle. Most companies build something powerful and immediately ship it to millions of users.
Anthropic built something so powerful they said "we're not releasing this publicly, it's too capable." Instead, they assembled a coalition of competitors to figure out how to handle it safely.
That decision alone should tell you everything you need to know about what we're dealing with.
Claude Mythos Preview isn't a security tool, a coding assistant, or a research platform. It's a general-purpose AI model that happens to be better at most intellectual tasks than most humans who specialise in those tasks.
Let me be specific about what "most capable ever built" actually means.
To put it bluntly: this model is better at coding than most senior developers, better at science than most PhDs, and better at security research than most hackers.
On GPQA Diamond, which tests graduate-level scientific reasoning, it scored 94.6%. This model understands science at a level that surpasses most PhDs in their own fields.
| Benchmark | What It Tests | Mythos Preview | Previous Gen (Opus 4.6) |
|---|---|---|---|
| SWE-bench Verified | Real-world software engineering | 93.9% | 80.8% |
| SWE-bench Pro | Advanced software engineering | 77.8% | 53.4% |
| GPQA Diamond | Graduate-level science questions | 94.6% | — |
| Terminal-Bench 2.0 | Advanced problem-solving | 82.0% | 65.4% |
| Humanity's Last Exam | The limits of AI reasoning | 64.7% | 53.1% |
| CyberGym | Cybersecurity capabilities | 83.1% | 66.6% |
| BrowseComp | Web research tasks | 86.9% | — |
On SWE-bench Verified, a benchmark that tests real-world software engineering problems, Mythos scored 93.9% compared to 80.8% for the previous generation (Anthropic Frontier Red Team, April 2026).
On the harder SWE-bench Pro benchmark, it hit 77.8% versus 53.4%. That's not incremental improvement. That's a generational leap.
Now, let's talk about the part that should make you sit up and pay attention.
When Anthropic's Frontier Red Team tested Mythos Preview's cybersecurity capabilities, they found zero-day vulnerabilities in every major operating system and browser they tested.
What the Red Team Found
Perhaps most concerning: Anthropic reported that researchers with no prior exploitation experience were able to create working exploits overnight using the model.
We've crossed a threshold where AI can outsmart decades of human expertise in specialised fields.
Perhaps most concerning: Anthropic reported that researchers with no prior exploitation experience were able to create working exploits overnight using the model.
We've crossed a threshold where AI can outsmart decades of human expertise in specialised fields.
A 27-year-old bug that countless expert eyes missed was found by a model trained on patterns. A 16-year-old vulnerability that survived years of security audits was discovered by software.
The defenders-versus-attackers dynamic is fundamentally shifting.
Think of it like the fuzzer revolution in the 2000s.
Automated testing tools found bugs faster than humans, which sounded scary until we realised it meant we could find and fix them before attackers did.
We're at that same inflection point, only the scale is orders of magnitude greater.
Forget the benchmarks for a moment. Let's talk about what you can actually do with AI at this capability level.
Automated code review, bug detection, refactoring, and feature development that used to take days now happens in hours.
Not because you're replacing developers, but because you're amplifying them.
According to a McKinsey study (2025), organisations using AI-augmented development reported 20-45% productivity gains in their engineering teams.
Market research, competitive intelligence, patent analysis, regulatory compliance reviews.
Tasks that used to require teams of analysts can now be done in a fraction of the time with greater thoroughness.
On the BrowseComp benchmark (web research tasks), Mythos scored 86.9% whilst using 4.9 times fewer tokens than comparison models.
It's not just better, it's more efficient.
Contracts, proposals, technical documentation, regulatory filings. The boring, time-consuming work that drains your team's capacity can now be automated whilst maintaining (or exceeding) quality.
For industries like financial services, legal, and healthcare, this alone could reclaim thousands of hours per year.
Imagine having an expert adviser who's read every relevant paper, regulation, case study, and market report in your industry, and can synthesise it all into actionable recommendations in seconds.
That's what a 94.6% score on graduate-level science questions actually means in practice.
The businesses that figure out how to integrate AI at this capability level into their operations will have an advantage that's difficult to overstate. The ones that don't will find themselves competing with organisations that operate at a fundamentally different speed.
At Aztech, we're seeing this play out in real time. Clients who embraced AI-augmented workflows six months ago are already operating differently, making faster decisions, handling more complexity with fewer bottlenecks, and freeing their teams to focus on work that actually moves the needle.
The gap between early adopters and everyone else is widening, and with each generation of AI capability, it becomes harder to close.
Project Glasswing brings together 12 launch partners that normally compete fiercely. They're now collaborating on a single initiative to secure the world's most critical software.
The partner organisations include Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Over 40 additional organisations building or maintaining critical infrastructure have also received access to the model for defensive security work.
Anthropic has backed Project Glasswing with $100 million in computing credits to help organisations scan for and patch vulnerabilities.
On top of that, $4 million in direct donations to open-source security organisations, specifically $2.5 million to Alpha-Omega and the Open Source Security Foundation (OpenSSF) via the Linux Foundation, plus $1.5 million to the Apache Software Foundation.
I have to give credit where it's due. In an industry that's been defined by "move fast and break things," Anthropic chose a different path.
They built the most capable model in the world and their first instinct was restraint, not a press release. That's significant.
Compare this to how previous model launches have gone.
We've watched companies race to ship increasingly powerful models with minimal safety testing, treating the general public as an unwitting beta test group.
Anthropic looked at what they'd built, recognised its potential for harm, and chose to assemble a coalition of the world's biggest technology companies to figure out responsible deployment before making it widely available.
This is how AI development should work. Safety and capability aren't opposing forces. They're complementary when you treat them seriously.
But here's where I'll be honest about my concerns. Look at the coalition list: AWS, Apple, Google, Microsoft, NVIDIA.
These are the biggest technology companies on the planet.
They're getting early access to a model that could fundamentally reshape their competitive position, whilst smaller businesses wait on the sidelines.
I understand the logic. You start with the organisations that maintain the critical infrastructure the rest of us depend on.
You give them the tools to find and patch the vulnerabilities that affect everyone. That makes sense from a security perspective.
But there's a real risk of widening the gap between the tech giants who get to shape these tools and the rest of the business world that eventually receives them.
The companies sitting at the Glasswing table aren't just patching code. They're building institutional knowledge about what the most capable AI in the world can do, and that knowledge compounds.
By the time smaller organisations get access, the big players will be a full cycle ahead.
That doesn't make Anthropic's decision wrong. It makes it imperfect, which is still vastly better than the alternative of dumping this capability onto the open market and hoping for the best.
But we should be eyes-open about what controlled access means for competitive dynamics in the medium term.
I'm optimistic. I want to be clear about that. But I'm the kind of optimistic that keeps one eye on the road and both hands on the wheel.
What excites me most is the compounding effect. AI that can find a 27-year-old bug can also find the cure that's been hiding in research data for decades.
AI that can generate 181 exploit chains can also generate 181 defensive strategies.
The same capability that makes this technology dangerous in the wrong hands makes it extraordinary in the right ones.
The Mythos announcement demonstrates something I've believed for a long time: the most powerful technology demands the most responsible stewardship.
Anthropic has shown that it's possible to push the boundaries of what AI can do whilst simultaneously taking seriously the question of what it should do.
That's a template the rest of the industry needs to follow.
But the optimism has to be earned, not assumed. The security findings are a sobering reminder that this technology doesn't just raise the ceiling for productivity and innovation.
It raises the floor for what's possible in terms of harm. Every organisation, government, and individual is going to need to recalibrate their understanding of risk in the coming months.
The businesses, institutions, and governments that engage with this technology thoughtfully and early will shape the future.
The ones that wait will be shaped by it.
7 April 2026 will be remembered as the day AI capability reached a level that forced the industry to rethink how these systems are released and governed. Anthropic didn't just build a more powerful model.
They demonstrated a more responsible approach to deployment.
At Aztech, we're already preparing.
We're building AI readiness into every conversation we have with clients, from cybersecurity posture assessments that account for AI-speed threats, to operational reviews that identify where AI augmentation can transform workflows.
This isn't a future consideration. It's a present reality, and we're helping organisations navigate it right now.
For your business, the choice is clear: engage with this technology now, or watch your competitors build advantages that become increasingly difficult to close.
The model may not be publicly available, but its implications are already reshaping every industry.
If you're reading this and thinking "I don't know where to start," that's exactly the right place to be.
Awareness is step one.
Step two is an honest assessment of where you are today.
Step three is building a plan to get where you need to be.
We'd be happy to help with steps two and three.
Ready to Assess Your AI Readiness?
Talk to our team about an AI audit for your organisation.
We'll help you identify the opportunities, address the security implications, and build a roadmap that fits your business.
Claude Mythos Preview is Anthropic's most capable AI model, announced on 7 April 2026. It is a general-purpose language model that outperforms previous AI systems across coding, science, reasoning, and cybersecurity benchmarks. Anthropic chose not to release it publicly due to its unprecedented capabilities, particularly in finding and exploiting software vulnerabilities.
Project Glasswing is an industry coalition launched by Anthropic to use Claude Mythos Preview for defensive cybersecurity.
It brings together 12 major technology companies, including AWS, Apple, Google, Microsoft, and NVIDIA, with $100 million in usage credits and $4 million in donations to open-source security organisations. Over 40 additional organisations maintaining critical infrastructure have also received access.
Anthropic determined that the model's ability to find and exploit software vulnerabilities at superhuman levels presented too significant a risk for general release.
During red team testing, it discovered zero-day vulnerabilities in every major operating system and web browser, and researchers with no prior exploitation experience were able to create working exploits overnight.
Even though you don't have access to Mythos, its existence changes your threat model.
The capabilities it demonstrates will eventually appear in other models, both commercial and open-source. Businesses should accelerate patching cadences, audit open-source dependencies, and review incident response plans for AI-speed threats.
Start with an AI audit: map your processes, identify where AI could have the biggest operational impact, and prioritise opportunities.
On security, accelerate patching, review your open-source supply chain, and update your incident response plans.
Build AI literacy across your leadership team, as the decisions made in the next 12 months will define your competitive position for years to come.
Access is limited to 12 Project Glasswing launch partners (AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks) plus over 40 additional organisations that build or maintain critical software infrastructure. The model is available via the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry for approved partners.
Claude Mythos Preview leads every major benchmark where comparisons are available. It scores 93.9% on SWE-bench Verified (vs 80.8% for its predecessor), 94.6% on GPQA Diamond (graduate-level science), and 82.0% on Terminal-Bench 2.0 (vs 65.4%). In cybersecurity testing, it produced 181 working Firefox exploits compared to just 2 from the previous generation. However, direct comparisons with unreleased competitor models are limited.
Founder & CEO of Aztech IT Solutions, a UK-based MSP established in 2006. With 19 years of experience in managed IT services, cybersecurity, and digital transformation, Sean helps organisations leverage technology for competitive advantage.
Connect on LinkedIn