Anthropic’s February and March have been characterized by an unprecedented pace of product development, with the AI research company consistently rolling out new features and capabilities for its Claude models. This aggressive release schedule, while demonstrating significant technological advancement, has also highlighted challenges in maintaining infrastructure stability. The company’s recent activities, including the accidental leak of information regarding its next-generation model, Claude Mythos, and the burgeoning adoption of its Model Context Protocol (MCP), paint a picture of a company pushing the boundaries of AI development at breakneck speed, while simultaneously grappling with the operational realities of rapid scaling.
Anthropic’s Unrelenting March: A Cascade of Claude Upgrades
For industry observers and practitioners alike, the past two months have been a testament to Anthropic’s intensified development cycle. The company has consistently delivered upgrades to its Claude models, with new features and enhancements appearing at an almost daily cadence. This rapid iteration is not merely incremental; it represents a significant acceleration in the company’s ability to ship new functionalities, a pace that veteran technology journalists find remarkable.
The major releases throughout February and March underscore this momentum. While specific dates for all updates were not universally publicized, the cumulative effect has been a substantial expansion of Claude’s capabilities. These advancements have touched upon core functionalities, including enhanced reasoning, improved coding assistance, expanded context windows, and refined safety protocols. The introduction of features such as updated code interpreter capabilities, advanced summarization tools, and more nuanced conversational agents reflects a strategic effort to broaden Claude’s applicability across various professional domains.
Beyond these headline announcements, a steady stream of smaller, yet significant, updates has refined the user experience and developer integration. Claude Code, for instance, has seen the addition of PowerShell support for Windows, alongside improvements to transcript search, deduplication of multimodal context, and more efficient idle-return prompt management. These incremental updates, often delivered through multiple point releases within a single week, contribute to a more robust and user-friendly developer ecosystem. Similarly, Anthropic’s collaborative workspace, Cowork, has benefited from the integration of plugin support and enhanced file management functionalities, further solidifying its utility for team-based AI workflows.
The New Stack has been at the forefront of reporting on these developments, providing in-depth coverage of key launches. Notably, their reporting on the multi-agent code review tool for Claude Code and the expansion of Claude Code to web and mobile platforms highlights Anthropic’s commitment to making its AI tools more accessible and powerful for developers.
The Double-Edged Sword of Rapid Deployment: Infrastructure Strain
This relentless pace of innovation, however, comes with a significant operational challenge: maintaining the stability and reliability of Anthropic’s underlying infrastructure. The sheer volume and frequency of new releases have placed considerable strain on the company’s servers, leading to multiple service disruptions throughout March. At least five notable outages were recorded within the month, including two extended periods of unavailability that occurred in the same week. As of Friday morning, reports indicated that the Opus 4.6 model was experiencing elevated error rates, while the Sonnet 4.6 model remained operational, underscoring the dynamic and sometimes fragile nature of the deployed systems.
This challenge is not unique to Anthropic. Major players in the AI landscape, including OpenAI and Google, have encountered similar growing pains as user demand surges and models become increasingly complex. However, Anthropic’s exceptionally aggressive release cycle magnifies the visibility of these infrastructure issues. The adage that "the best model is useless if it’s not available" holds true. Developers and businesses seeking to integrate AI into their workflows rely on consistent uptime and predictable performance. Frequent outages, even if temporary, can erode trust and lead to the adoption of fallback strategies. These workarounds, initially intended as temporary measures, can often become permanent fixtures in system architecture, hindering the full adoption of advanced AI capabilities. Reliability, therefore, is emerging as a critical product feature for AI providers, and for Anthropic, it may represent the most pressing area requiring immediate attention and robust delivery.
Claude Mythos: A Powerful Next-Gen Model Revealed by Data Leak
In a development that underscored the rapid evolution within Anthropic, Fortune reported on Thursday that the company is currently testing a new, next-generation AI model internally codenamed "Capybara," publicly revealed as Claude Mythos. Anthropic describes this upcoming model as a "step change" in capability, representing the "most capable we’ve built to date."
The premature revelation of Claude Mythos was attributed to an unintentional data leak. Anthropic acknowledged that "human error" during a content management system (CMS) configuration process resulted in approximately 3,000 unpublished assets, including a draft blog post detailing the model, being left in a publicly accessible data cache. Security researchers discovered the exposed information, and Fortune was able to review it before Anthropic secured the data.
The claimed capabilities of Claude Mythos are substantial and potentially game-changing. According to the leaked documentation, the model exhibits dramatically improved performance on benchmarks related to coding, academic reasoning, and cybersecurity. Anthropic’s own assessment, as detailed in the leaked materials, positions Claude Mythos as "far ahead of any other AI model in cyber capabilities." The company further cautioned that the model "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders," a candid warning from the developers themselves about the power they are unleashing.
Anthropic has stated that Claude Mythos is currently undergoing testing with a select group of early access customers and that the company is adopting a deliberate approach to its public release. This controlled rollout, especially in light of the model’s advanced cybersecurity capabilities, suggests a strategic effort to manage the potential risks associated with such powerful technology. The confluence of aggressive feature shipping, infrastructure challenges, and the imminent arrival of a significantly more advanced model positions Anthropic at a complex juncture, attempting to balance rapid innovation with operational stability and responsible deployment.
Model Context Protocol Surpasses 97 Million Installs, Establishing Industry Standard
Anthropic’s Model Context Protocol (MCP) has achieved a significant milestone, crossing 97 million monthly SDK downloads in March. This remarkable growth trajectory, from approximately 2 million downloads at its launch in November 2025, represents a staggering 4,750% increase in just 16 months. The MCP ecosystem has expanded to encompass over 5,800 community and enterprise servers, integrating with a wide array of platforms including databases, CRMs, cloud providers, and developer tools.
The broad adoption of MCP signifies a critical shift in its status from an Anthropic-specific protocol to an industry standard. This transition was further cemented when OpenAI committed to supporting MCP last year. In a move to formalize this industry-wide adoption, Anthropic donated MCP to the Agentic AI Foundation in December. This foundation, co-founded with Block and OpenAI, operates under the umbrella of the Linux Foundation, ensuring open governance and collaborative development.
The New Stack has provided extensive coverage of MCP’s evolution. Articles examining the reasons behind MCP’s widespread adoption and its 2026 roadmap have delved into the ongoing efforts to address production-readiness gaps. These include enhancements in authentication, observability, and large-scale server management. While the protocol is clearly gaining traction, the journey to full production readiness for all use cases remains a significant undertaking, as highlighted in previous reporting. For developers building agentic workflows, MCP has become the de facto standard. The crucial question now is whether the surrounding tooling and infrastructure can scale effectively to meet the burgeoning demand.
Anthropic, alongside other key stakeholders, will be participating in the MCP Dev Summit in New York City from April 2nd to 3rd. This event promises to be a crucial forum for discussing the protocol’s future development and addressing the challenges of integrating it into production environments. Alex Wilhelm, publisher of the Cautious Optimism newsletter, will be present, conducting interviews for The New Stack, further underscoring the industry’s keen interest in MCP’s progress.
The AI Czar Departs, Leaving a Void in Policy Leadership
David Sacks, who served as a special government employee advising on artificial intelligence and cryptocurrency policy, announced on Thursday that he has concluded his 130-day term and is stepping down from his role as the Trump administration’s "AI and crypto czar." He will now co-chair the President’s Council of Advisors on Science and Technology (PCAST) alongside Michael Kratsios. This transition broadens Sacks’s portfolio but potentially diminishes his direct influence on AI policy formulation.
During his tenure as AI czar, Sacks held a direct line to President Trump and played a role in shaping policy initiatives. In his new capacity as PCAST co-chair, his role will shift towards making recommendations rather than direct policy implementation. Reports from Axios indicate that the White House does not intend to appoint a new AI czar, leaving a significant leadership vacancy in Washington’s AI policy landscape.
Prior to his departure, Sacks expressed optimism regarding the potential for bipartisan AI legislation. He suggested to Bloomberg that Congress could pass such legislation within months, outlining a framework that would pair child safety measures with federal preemption of state-level AI laws. Sacks noted a positive reception on Capitol Hill, stating, "This is an area where I think we’re willing and happy to work with Democrats." The durability of this bipartisan momentum following his departure remains a key question. AI policy may represent one of the few genuinely bipartisan issues currently, but without a dedicated champion, such momentum can dissipate.
For organizations actively adopting AI, the evolving regulatory landscape is of paramount importance. Federal preemption of state laws would significantly streamline compliance efforts, alleviating a major point of friction for enterprises deploying AI solutions across the United States. If Sacks’s projection of impending federal legislation proves accurate, this development warrants close attention from the business and technology sectors. The potential for a unified federal approach to AI regulation could profoundly impact the speed and scope of AI adoption nationwide.