Back to Blog

    Anthropic's Claude Mythos Leak: A Step Change in AI — And Why Your Growth Data Strategy Matters More Than Ever

    berner-setterwallMarch 27, 20266 min read

    Yesterday, Fortune broke the news that Anthropic is testing a new AI model that the company itself calls "a step change" in capabilities. The model, codenamed Claude Mythos (also referred to internally as "Capybara"), was revealed through an accidental data leak — draft blog posts left in a publicly accessible data cache.

    This isn't just another incremental model update. Anthropic's own words: "the most capable we've built to date." As a company building on Claude every day, we've been paying close attention. Here's what we know, what we think it means, and why it matters for anyone doing growth work.

    What We Know So Far

    The details come from Anthropic's own draft blog post, inadvertently made public, and confirmed by the company in a statement to Fortune:

    • A new model tier called "Capybara" — larger and more intelligent than Opus, which was previously their most powerful tier. This sits above Opus, Sonnet, and Haiku in the lineup.
    • Dramatically higher scores on software coding, academic reasoning, and cybersecurity benchmarks compared to Claude Opus 4.6 (the current best).
    • Cautious rollout — currently being trialed with early access customers. The model is expensive to run and not yet ready for general release.
    • Anthropic considers it unprecedented in terms of capability and has flagged cybersecurity risks, suggesting the reasoning and tool-use improvements are substantial.

    The fact that Anthropic is treating this as a "step change" rather than an iteration is significant. In our experience building on Claude, each major capability jump has unlocked entirely new categories of what AI agents can do autonomously.

    Why This Matters for Growth Teams

    Let's be specific about what "step change in reasoning and coding" means for growth work.

    Better Reasoning = Better Growth Ideation

    Growth experimentation is fundamentally a reasoning problem. You have data from multiple channels, historical experiment results, competitive signals, and business context. The quality of your next experiment idea depends on how well you can synthesize all of that.

    Today's models are already good at this. Claude Opus 4.6 can look at your Google Ads data, cross-reference it with SEO trends, and suggest experiments you hadn't considered. But there are limits — complex multi-step reasoning chains sometimes lose the thread, and connecting dots across very different data sources can be hit or miss.

    A genuine step change in reasoning means:

    • Deeper cross-channel analysis — finding non-obvious connections between paid, organic, and product data
    • More sophisticated experiment design — not just "test this headline" but "here's a multi-variant strategy based on your cohort behavior patterns"
    • Better causal reasoning — moving from "this metric changed" to "here's why, and here's what to do about it"

    Better Coding = Better Data Analysis

    The coding improvements matter too. When an AI agent analyzes your growth data, it's writing SQL, processing results, and building on previous queries. Better coding capability means:

    • More complex analytical queries executed correctly on the first try
    • Better handling of messy, real-world data schemas
    • More reliable multi-step analysis pipelines

    We've seen this progression firsthand. Each Claude model improvement has directly translated to fewer failed queries and more insightful analysis in Cogny.

    The Real Question: Will Your Data Be Ready?

    Here's the thing that most teams miss when a new model drops: the model is only as good as the data and context it has to work with.

    When Claude Mythos becomes generally available, the teams that benefit most won't be the ones who sign up fastest. They'll be the ones who have been systematically collecting and structuring their growth data.

    Think about what a step-change reasoning model could do if it had access to:

    • Your complete experiment history — every A/B test, every landing page variant, every bid strategy change, with results and learnings attached
    • Cross-channel performance data — not in separate dashboards, but unified in a queryable warehouse
    • Institutional knowledge — what you've tried before, what worked, what didn't, and why
    • Competitive context — how your performance compares to market benchmarks

    Now compare that to a team that's starting from scratch, with data scattered across platform UIs and tribal knowledge locked in people's heads. Same model, vastly different outcomes.

    Why We're Bullish on Structured Growth Data

    This is exactly what we've been building Cogny around. Not because we predicted the Mythos leak (we didn't), but because the pattern has been clear for a while:

    AI models keep getting better. The bottleneck is shifting from model capability to data quality and context.

    Every growth experiment you run, every result you record, every piece of context you add to your data warehouse — it's all compound interest. Each new model generation makes that accumulated data more valuable, not less.

    Here's what we recommend, regardless of when Mythos ships:

    1. Connect your data sources now. Get your Google Ads, Meta, LinkedIn, Search Console, and analytics data flowing into a unified warehouse. The setup cost is a one-time investment; the value compounds with every model improvement.

    2. Document your experiments. Not in Notion docs that nobody reads. In a structured format that an AI agent can actually reason over. What did you test? What happened? What did you learn?

    3. Build organizational context. Your industry, your competitors, your strategy, your constraints — this is the context that turns generic AI analysis into actionable growth insights.

    4. Start using AI-assisted analysis today. Don't wait for Mythos. The teams that are already working with AI agents for growth analysis are building intuition about what works and what doesn't. That experience compounds too.

    What We're Watching

    A few things we're keeping an eye on as more details emerge:

    • Pricing and availability. Anthropic has signaled that Capybara-tier models will be more expensive than Opus. We'll need to see the actual pricing to understand the cost-performance tradeoff for production workloads.
    • Tool use improvements. The coding benchmarks are promising, but what matters for us is how well the model handles complex, multi-step tool use — querying BigQuery, processing results, and building on findings iteratively.
    • Context window. Bigger reasoning capability combined with larger context windows would be transformative for growth analysis, where you often need to hold an entire quarter's worth of data in context.
    • Rollout timeline. "Early access customers" today, but when does GA happen? We'll be among the first to test and share results.

    The Bottom Line

    The Claude Mythos leak confirms what we've been betting on: AI capabilities for growth and marketing analysis are about to take another major leap forward.

    But capability without context is just potential. The teams that win aren't the ones with access to the best model — everyone will have that eventually. The teams that win are the ones whose data, experiments, and institutional knowledge are structured and ready for whatever the next model can do.

    Start collecting. Start structuring. Start compounding.

    The best time to prepare your growth data for the next AI step change was six months ago. The second best time is now.