Anthropic announced a preview of Claude Mythos this week and in the same breath said it would not be making the model generally available. The reason: the model is, by their own assessment, too capable at finding and exploiting security vulnerabilities to release responsibly. In evaluation, Mythos autonomously identified thousands of zero-day flaws across every major operating system and web browser, including a bug in OpenBSD that had gone undetected for 27 years.

Rather than shelve the model, Anthropic launched Project Glasswing, a restricted access program giving a cybersecurity coalition including AWS, Apple, Cisco, CrowdStrike, Google, JPMorgan Chase, Microsoft, and Nvidia access to Mythos Preview for defensive security work. Roughly 40 additional organizations responsible for critical software infrastructure were also granted access.

For marketers, the implications are less about cybersecurity and more about what this signals for AI capability timelines. Anthropic is now operating a model it considers too powerful to ship broadly. The same coding and reasoning depth that makes Mythos a security risk is the same architecture that will eventually reach Claude Code and downstream agentic tools.

Meta shipped Muse Spark this week, the first model from Meta Superintelligence Labs under Alexandr Wang. It is also the first Meta model that is not open-weight, a sharp break from the Llama strategy. Muse Spark scores 52 on the Artificial Analysis Intelligence Index, placing it behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6 but ahead of the Llama 4 family by a wide margin.

The model powers the Meta AI app immediately, with rollout to WhatsApp, Instagram, Facebook, and Messenger in the next few weeks. A private API preview is available to select partners. Meta's shopping mode, which combines the model with user behavior signals, is the clearest signal of where Zuckerberg is taking this. Meta stock jumped nearly 10% over five trading days following the announcement.

Alibaba's Qwen 3.6-Plus launched via the Alibaba Cloud API this week with native 1M token context support, making it one of the most capable open-access models for long-horizon agentic tasks. The model introduces a preserve_thinking parameter that retains reasoning context across message turns, reducing redundant token use in multi-step workflows.

Qwen 3.6-Plus is compatible with Claude Code via the Anthropic API protocol, OpenClaw, and Qwen Code. For marketing tech teams building automation pipelines, the 1M context window and agentic compatibility make this worth benchmarking against proprietary alternatives at a fraction of the cost.

OpenAI's ChatGPT ad pilot crossed $100M in annualized revenue within six weeks of launch, with over 600 advertisers in the managed program. Self-serve access opens to a broader market this month. Internal projections shared with investors target $2.5B in ad revenue this year, scaling to $100B annually by 2030 assuming 2.75B weekly users.

Ads appear below responses, clearly labeled and separated from ChatGPT's answers. OpenAI says advertisers do not receive user chat data and reports no measurable impact on consumer trust metrics. Anthropic responded with a formal statement that Claude products will remain ad-free, calling advertising incentives structurally incompatible with a genuinely helpful AI assistant.

Google's Newfront 2026 presentation revealed full Gemini integration across Display and Video 360, with AI-curated media packages that proactively plan where budgets should go before campaigns go live. New formats include YouTube Creator Takeovers, Pause Ads, and Confidential Publisher Match, which connects first-party advertiser data with CTV streaming signals inside Trusted Execution Environments. Google reported an 80% revenue lift for one retailer after enabling AI Max campaigns, which eliminate keyword targeting entirely and use Gemini to match landing pages with user intent.

For media buyers, the shift is structural: Google is moving from a platform you configure to one that configures itself. The human role shifts from managing bids and audiences to supplying the best possible inputs and strategy.

Claude Cowork reached general availability on Mac and Windows this week, bringing computer use to Pro and Max subscribers with no setup required. Claude can now open files, click, scroll, and navigate your screen to perform tasks autonomously. Enterprise plans get role-based access controls, OpenTelemetry monitoring, and usage analytics.

Anthropic also launched an advisor/executor agent pairing for long-horizon tasks. A high-intelligence advisor model provides strategic guidance mid-generation while a faster executor model handles the bulk of token generation, delivering near-advisor-quality output at significantly lower cost. This is the architecture that makes complex multi-step marketing automation practical at scale.