New Peak-Hour Session Limits Take Effect — What Changes Today
With the two-week 2x off-peak usage boost now expired, Anthropic has quietly shifted to a new capacity management model. Free, Pro, Max 5x, and Max 20x subscribers now face tighter session limits during peak hours — roughly 5am–11am PT and 1pm–7pm GMT — where a five-hour session window can be exhausted faster than before. Anthropic says overall weekly limits are unchanged; the difference is distribution across the day.
The Register and gHacks both confirmed the change. Separately, MacRumors is tracking a bug report from Claude Code users who say their rate limits are draining dramatically faster than expected, suggesting a possible accounting error in the new limit logic. Anthropic has not issued an official statement on the Claude Code issue but community threads on Reddit and Discord are active.
Structured Outputs Now GA — No Beta Header Required
Anthropic quietly graduated structured outputs from beta to general availability on the Claude API. Developers on Claude Sonnet 4.5, Opus 4.5, and Haiku 4.5 no longer need to pass a beta header to enable schema-constrained responses. The GA release also brings expanded schema support, lower grammar compilation latency, and a simplified integration path. Data residency controls are also live, letting API customers specify whether inference runs in US-only infrastructure at a 1.1x price premium.
MCP Hits 97M Monthly Downloads — Anthropic Hands Governance to Linux Foundation
The Model Context Protocol has reached 97 million monthly SDK downloads — up from 2 million at launch in November 2024, a 4,750% increase in 16 months. For context: npm’s React package took roughly three years to hit comparable scale. The protocol now underpins 5,800+ community and enterprise server integrations spanning databases, CRMs, cloud providers, and developer tools.
Alongside the milestone, Anthropic donated MCP governance to the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation co-founded with OpenAI, Block, Google, Microsoft, AWS, and Cloudflare. The move signals that MCP is no longer an Anthropic-specific protocol — it’s becoming industry infrastructure, with every major AI provider now supporting it as a standard.
Claude Code Rate Limit Bug Reported — Users Seeing Rapid Drain
A growing number of Claude Code users on Team and Pro plans are reporting that their session limits are being consumed far faster than expected — sometimes exhausting a five-hour window in under two hours. MacRumors was first to report the issue, with users pointing to the rollover from the now-expired 2x boost promotion as a possible trigger. Some developers have reported the issue to Anthropic support. No official fix or acknowledgment has been posted to the Claude status page as of this writing.
Accenture Launches Cyber.AI — Claude Becomes the Brain of Enterprise Security Ops
Accenture introduced Cyber.AI this week, a new security operations platform that puts Claude at its reasoning core. The system synthesizes security data, identifies vulnerabilities, and provides contextual insights across the full security lifecycle. Accenture deployed it internally first: scan turnaround on 1,600 applications dropped from three-to-five days to under one hour, and security testing coverage expanded from 10% to over 80% across 500,000+ APIs.
The platform builds on the multi-year Anthropic-Accenture partnership announced in December 2025, which includes training roughly 30,000 Accenture professionals on Claude. Cyber.AI adds an enterprise governance layer — called Agent Shield — that enforces organizational policies on every agent action. UBS analysts called the partnership a signal that strengthens the AI thesis for both companies. The timing is notable given the Claude Mythos leak’s focus on cybersecurity capabilities.
Mythos Capybara Tier: 10 Trillion Parameters and Training Complete
New reporting from Geeky Gadgets and SiliconAngle adds detail to the Claude Mythos story. The model reportedly uses a 10-trillion parameter architecture — far beyond the scale of current Opus models — and represents a new “Capybara” product tier that would sit above Opus in Anthropic’s lineup and carry a higher price point. Most significantly, Anthropic has reportedly completed training, which means the model’s capabilities are now fixed and deployment decisions are the remaining variable.
Anthropic has said no public release is planned in the near future, citing the unprecedented cybersecurity risks the model poses. But industry analysts note that “near future” is vague, and the IPO timeline reported by Bloomberg adds commercial pressure to accelerate. The Coindesk analysis notes the leak itself sent software stocks briefly lower before recovering.
The New Stack: Anthropic’s “Madcap March” — 14+ Launches, 5 Outages, 1 Leaked Model
The New Stack published a sweeping recap of Anthropic’s March 2026, calling it a month unlike any in the company’s history. The tally: more than 14 product launches (including Claude Code auto mode, computer use for Cowork, Claude for Excel and PowerPoint, web search GA, voice mode, and structured outputs), five service outages of varying severity, a landmark court victory over the Pentagon, a Bloomberg-reported IPO, and a data leak that revealed the company’s most powerful model to date. All in 28 days.
The Usage Limit Signal: Anthropic Is Running Hot
The shift to peak-hour session rationing isn’t just a pricing move — it’s an infrastructure signal. When a company with Anthropic’s compute budget starts throttling during business hours, it tells you demand is outpacing GPU supply. That’s a good problem to have heading into an IPO, but it’s also a real friction point for the developer community that Anthropic depends on for ecosystem loyalty.
The MCP donation to the Linux Foundation is a savvy counterbalance to that friction. By giving MCP to open governance, Anthropic is making a bet that ecosystem standardization serves them better than lock-in. It also insulates the protocol against the risk that a competitor forks it if they feel Anthropic controls too much of the agentic AI stack. Meanwhile, the Accenture Cyber.AI deal shows how Anthropic’s enterprise strategy is maturing: less about selling API tokens and more about being the reasoning engine inside industry-specific platforms. That’s a higher-margin play — and it’s one that will show up nicely in an IPO prospectus.