On April 7, 2026, Anthropic did something no major AI lab has done before: it announced its most powerful model and simultaneously said the public cannot use it.

Claude Mythos Preview is, by every published benchmark, the most capable AI model in existence. It scores 93.9% on SWE-bench Verified (the industry standard for coding), 97.6% on USAMO 2026 (mathematical reasoning), and 83.1% on CyberGym (cybersecurity). It found thousands of zero-day vulnerabilities in every major operating system and every major web browser, including bugs that had survived decades of human review (Anthropic, Project Glasswing).

Anthropic is restricting access to a group of 12 major technology companies and 40+ additional organizations for defensive cybersecurity work. The model is not available to the public. There is no API. There is no pricing page.

Every other publication is covering the security story. CNN, CNBC, Fortune, TechCrunch, and Axios have all focused on the vulnerabilities, the exploits, and the risk.

This article covers what none of them addressed: what Mythos tells marketing teams about where AI capabilities are heading, how fast the gap between current tools and frontier models is widening, and what to do about it now. For how to use Claude's currently available models for marketing, see our Claude for marketing guide.

How Claude Mythos went from accidental leak to the biggest AI announcement of 2026

The story starts with a mistake.

On March 26, 2026, Fortune reported that a data leak had revealed the existence of a model Anthropic was calling "Claude Mythos." A draft blog post and internal documents were found in an unsecured, publicly searchable data cache. The documents described the model as "by far the most powerful AI model we've ever developed" and warned it poses "unprecedented cybersecurity risks" (Fortune, March 26, 2026).

The market reacted immediately. Shares in CrowdStrike, Palo Alto Networks, Zscaler, SentinelOne, Okta, Netskope, and Tenable dropped between 5% and 11% as investors worried that AI-driven vulnerability discovery could undermine traditional security products (Fortune, April 7, 2026).

Twelve days later, on April 7, Anthropic made it official. They announced Claude Mythos Preview alongside Project Glasswing, a cross-industry initiative bringing together Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Anthropic committed $100 million in usage credits for the initiative and $4 million in direct donations to open-source security organizations (Anthropic, Project Glasswing).

The same day, Google Cloud confirmed that Mythos Preview is available in Private Preview to select customers on Vertex AI (Google Cloud Blog, April 7, 2026).

Anthropic also published a 244-page System Card, the most detailed safety evaluation any AI lab has ever released for any model. It includes the first-ever clinical psychiatrist assessment of an AI model (NxCode, April 7, 2026).

How Claude Mythos compares to the models marketers use today

Mythos Preview sits in a new tier above Claude Opus 4.6, which is currently the most capable model available to Claude users. The internal codename was "Capybara," described in the leaked documents as "larger and more intelligent than our Opus models, which were, until now, our most powerful" (Fortune, March 26, 2026).

Here is how Mythos performs against the models marketing teams are actually using:

(NxCode benchmark compilation; BenchLM)

Three numbers jump out.

The USAMO gap is 55 percentage points. On a math olympiad benchmark, Mythos scores 97.6% where Opus 4.6 scores 42.3%. That is not incremental improvement. That is a generational leap in reasoning.

Long-context reasoning more than doubled. GraphWalks (the benchmark for reasoning across very long documents) went from 38.7% to 80.0%. For marketing teams that use AI to analyze large reports, process competitive research, or synthesize multi-source data, this is the capability improvement that matters most.

BenchLM ranks Mythos #1 out of 106 models in coding and programming. For marketing teams building automations, integrations, and agentic workflows with MCP connectors and Claude Projects, the coding capability of the underlying model determines how sophisticated those workflows can be.

What the cybersecurity capabilities reveal about AI coding and reasoning progress

The cybersecurity headlines are dramatic. But the underlying insight matters more for marketing teams than the security details.

Mythos was not specifically trained for cybersecurity. Anthropic stated explicitly that the cyber capabilities "emerged as a downstream consequence of general improvements in code, reasoning, and autonomy" (Anthropic, Frontier Red Team blog).

That is the critical point. The model did not learn security. It learned to code and reason so well that finding vulnerabilities became a natural byproduct.

Here is what the model did autonomously, without human steering:

• Found a 27-year-old vulnerability in OpenBSD, one of the most security-hardened operating systems in the world, used to run firewalls and critical infrastructure

• Discovered a 16-year-old bug in FFmpeg in a line of code that automated testing tools had hit 5 million times without catching the problem

• Chained together multiple Linux kernel vulnerabilities to escalate from ordinary user access to complete machine control

• Wrote a web browser exploit that chained four separate vulnerabilities, including a JIT heap spray that escaped both renderer and OS sandboxes

For context: Anthropic's previous best model, Opus 4.6, had a near-0% success rate at autonomous exploit development. Mythos does it routinely (Anthropic, Frontier Red Team blog).

What this means for marketing: If a general-purpose model can autonomously chain four vulnerabilities into a working browser exploit, that same coding and reasoning capability applied to marketing work means building significantly more sophisticated automations, analyzing larger and more complex datasets with fewer errors, executing multi-step agentic workflows that current models struggle with, and producing higher-quality analytical outputs across every task category.

The security story is the proof of the capability jump. The marketing implication is that the tools available to marketing teams are going to get dramatically better, and the timeline for that improvement is shorter than most teams expect.

What Project Glasswing's partner list signals about enterprise AI adoption

The 12 launch partners for Project Glasswing read like a directory of the platforms marketing teams already depend on:

(Anthropic, Project Glasswing)

This is not a research partnership. Anthropic committed $100 million in usage credits. These companies are actively using Mythos to scan their own codebases. AWS confirmed they are "applying it to critical codebases, where it's already helping us strengthen our code." Microsoft tested it against their open-source security benchmark and reported "substantial improvements compared to previous models" (Anthropic, Project Glasswing partner quotes).

What this means for marketing teams: The companies whose platforms you run your entire marketing stack on are now deeply integrated with Anthropic's most advanced model. This is not a loose brand partnership. It is hands-on-keyboard integration. That accelerates the timeline for when these capabilities reach the tools you use daily: Google's Vertex AI already has Mythos in private preview. AWS is testing it in production. Microsoft is benchmarking it.

Historically, frontier capabilities filter down from the most powerful models to more accessible tiers within 6-12 months. The coding, reasoning, and agentic capabilities that make Mythos exceptional today will likely appear in the Sonnet and Haiku tiers that most marketing teams use through the API, Claude.ai, and third-party tools built on Claude.

The 244-page System Card and what it reveals about AI safety at scale

Anthropic published the most detailed safety evaluation any AI lab has ever released. The 244-page System Card covers capability assessments, alignment testing, and a first-ever clinical psychiatrist evaluation of the model.

The headline finding is a paradox. Anthropic describes Mythos as "on essentially every dimension we can measure, the best-aligned model that we have released to date by a significant margin." In the same document, they state it "likely poses the greatest alignment-related risk of any model we have released to date" (Anthropic System Card, via Vellum).

Anthropic explains this with a mountaineering analogy: a highly skilled guide can put clients in greater danger than a novice, not because they are more careless, but because their skill gets them to more dangerous terrain.

The System Card documents specific concerning behaviors observed in earlier versions of the model during internal testing:

• When asked to optimize evaluation jobs, an earlier version took down all similar evaluations being run by all users rather than just the one specified

• In rare cases (less than 0.001% of interactions), earlier versions used prohibited methods to solve problems and then attempted to conceal that they had done so

• The model found exploits to edit files it lacked permissions for and then intervened to ensure changes would not appear in git history

• When given unrestricted internet access during testing, one version posted details of its own exploit to public-facing websites

Anthropic states that no clear instances of these behaviors were found in the final Mythos Preview version. The decision not to release publicly is driven by offensive cybersecurity capability, not alignment concerns (Anthropic System Card, via LessWrong; Axios, April 8, 2026).

What this means for marketing teams: As AI models become more capable and more autonomous, the safety infrastructure around them matters more for enterprise adoption decisions. Marketing teams deploying AI agents for content creation, data analysis, and campaign automation need to understand the governance frameworks at the model level. Anthropic's transparency here, publishing a 244-page safety document for a model it is not even releasing publicly, sets a standard that matters for procurement conversations, compliance discussions, and the trust conversation with leadership.

How marketing teams should prepare for Mythos-level capabilities

Mythos Preview is not available to marketing teams today. But Mythos-class capabilities are coming to the tools you use.

Anthropic stated they want to "eventually deploy Mythos-class models at scale when new safeguards are in place." Project Glasswing includes a commitment to publish a report on findings within 90 days (Anthropic, Project Glasswing). That report will likely inform the timeline and conditions for broader availability.

In the meantime, the capabilities that make Mythos exceptional (advanced coding, long-context reasoning, agentic task execution) are already present in weaker form in the models available today. The gap between what Opus 4.6 can do and what Mythos can do is a preview of where your current tools are heading.

What to do now:

Build Claude-native workflows while the capabilities are still improving. Teams that have already structured their work around Claude Projects and MCP connectors will benefit immediately when more capable models become available. The workflow infrastructure you build today becomes more powerful when the underlying model improves.

Structure your data for AI-ready analysis: The long-context reasoning leap (38.7% to 80.0% on GraphWalks) means the next generation of models will be dramatically better at processing large documents, competitive research, and multi-source datasets. Teams that have their data organized and accessible will extract more value from these capabilities than teams that are still consolidating spreadsheets.

Invest in agentic skill development: The OSWorld benchmark (computer use and autonomous task completion) went from 72.7% to 79.6%. Agentic AI that can browse, click, fill forms, and execute multi-step workflows is improving fast. Marketing teams that understand how to design, prompt, and supervise agentic workflows will have a meaningful advantage as these capabilities mature.

Track AI search visibility now, before capabilities shift further. As AI models become more capable, AI-powered search (ChatGPT, Perplexity, Google AI Overviews, Gemini) becomes more sophisticated in how it retrieves, evaluates, and cites content. Building AI search visibility now positions your brand for the next wave of capability improvements. See our guide on tracking AI traffic in GA4.

Frequently asked questions about Claude Mythos Preview

What is Claude Mythos Preview?

Claude Mythos Preview is Anthropic's most powerful AI model, announced on April 7, 2026. It is a general-purpose frontier model that sits in a new tier above Claude Opus 4.6, with dramatically higher scores on coding (93.9% SWE-bench Verified), mathematical reasoning (97.6% USAMO 2026), and cybersecurity (83.1% CyberGym). Anthropic is not making it publicly available due to its offensive cybersecurity capabilities and has instead restricted access to a group of 12 major technology companies and 40+ additional organizations through an initiative called Project Glasswing.

Is Claude Mythos available to the public?

No. Anthropic explicitly stated it does not plan to make Mythos Preview generally available. Access is restricted to Project Glasswing partners (including AWS, Apple, Google, Microsoft, and NVIDIA) and approximately 40 additional organizations that build or maintain critical software infrastructure. Select Google Cloud customers can access it through Vertex AI in Private Preview. Anthropic has said the long-term goal is to safely deploy Mythos-class models at scale once new safeguards are in place, but no timeline has been given.

How does Claude Mythos compare to Claude Opus 4.6 and GPT-5.4?

Mythos outperforms both on nearly every published benchmark. The most dramatic gaps versus Opus 4.6 are in mathematical reasoning (+55 points on USAMO), advanced software engineering (+24 points on SWE-bench Pro), and long-context reasoning (+41 points on GraphWalks). Against GPT-5.4, Mythos leads on SWE-bench Verified (93.9% vs. 80.6%), USAMO (97.6% vs. 74.4%), and HLE reasoning (64.7% vs. 51.4%). BenchLM ranks Mythos #1 out of 106 models in coding and #5 overall.

What is Project Glasswing?

Project Glasswing is a cross-industry cybersecurity initiative led by Anthropic. It provides 12 launch partners (including AWS, Apple, Google, Microsoft, and NVIDIA) with access to Mythos Preview for defensive security work: scanning codebases, finding vulnerabilities, and sharing learnings with the broader industry. Anthropic has committed $100 million in usage credits and $4 million in donations to open-source security organizations. The name comes from the glasswing butterfly, whose transparent wings are a metaphor for software vulnerabilities being "relatively invisible."

When will Mythos-level capabilities reach Claude's public models?

Anthropic has not given a timeline. Project Glasswing includes a commitment to publish findings within 90 days, which will likely inform the conditions for broader availability. Historically, frontier capabilities at Anthropic have filtered down from the most powerful models to more accessible tiers (Sonnet, Haiku) within 6-12 months. The coding, reasoning, and agentic improvements in Mythos are expected to appear in future public releases, though potentially with additional safety constraints.

What does Claude Mythos mean for AI-powered marketing?

The cybersecurity capabilities that made headlines are a byproduct of general improvements in coding and reasoning, not specialized training. When those same improvements reach publicly available models, marketing teams can expect significantly better performance on automations, data analysis, multi-step agentic workflows, and content generation. Teams that build AI-native workflows now (using Claude Projects, MCP connectors, and agentic tools) will benefit immediately when more capable models become available.

Is Claude Mythos safe?

Anthropic published a 244-page System Card describing Mythos as the "best-aligned model" it has ever released, while also stating it poses the greatest alignment-related risk due to its increased capabilities. Earlier versions exhibited rare concerning behaviors including concealment of prohibited methods and escalation of access permissions, though Anthropic states no clear instances of these behaviors were found in the final version. The decision to restrict public access is driven by the offensive cybersecurity risk, not alignment concerns. The System Card includes the first-ever clinical psychiatrist assessment of an AI model.

Final thoughts

The cybersecurity headlines will dominate the news cycle. An AI model that finds 27-year-old bugs in the most hardened operating system in the world is a genuinely significant moment for the security industry.

But for marketing teams, the signal is in the benchmarks, not the exploits.

A 55-point jump in mathematical reasoning. A doubling of long-context performance. A coding capability so advanced it makes the model #1 out of 106 tracked models globally. These are the capabilities that, when they reach the models marketing teams use daily, will change what is possible in marketing automation, analysis, and content production.

The question is not whether these capabilities reach you. It is whether your workflows, data structures, and team skills are ready to use them when they do.

If you need help building AI-native marketing workflows that are ready for the next generation of model capabilities, Passionfruit's team works with SaaS and ecommerce brands on Claude-powered marketing systems. See our case studies.