Anthropic hits capacity limits
Anthropic is facing unexpected capacity limits and periodic outages for flagship Claude models, prompting tweaks to usage limits as demand surges — and the company is reportedly testing a more powerful internal model, 'Mythos', with select clients after leaks. That combination of strain and clandestine rollouts underscores uptime risks when you depend on third‑party LLMs. (theregister.com, trendingtopics.eu)
Anthropic has redefined its five‑hour session allowance so that during peak windows (05:00–11:00 PT / 13:00–19:00 GMT) a five‑hour allocation can be consumed faster than clock time, while the company says weekly totals remain unchanged and roughly 7% of users will now hit session limits they previously wouldn’t. (theregister.com) Subscription tiers and published prices for Claude remain Free, Pro ($20/month), Max 5x ($100/month) and Max 20x ($200/month), and Anthropic ties hourly/session access to token consumption without disclosing the exact token-to‑time mapping. (theregister.com) Independent outage trackers showed a surge in incident reports this week, with IsDown recording about 820 user reports in a 24‑hour window and StatusGator logging 567 user‑submitted reports as outages clustered around March 25–27. (isdown.app) Anthropic briefly boosted off‑peak capacity by doubling usage allowances outside peak hours in a promo running from March 13 through March 27 (some customer notices extended that program through March 28), a stopgap measure the company said was to address rapid demand growth. (xda-developers.com) A content‑management misconfiguration left nearly 3,000 unpublished Anthropic assets publicly searchable in late March, a cache that outside researchers and journalists reviewed before Anthropic restricted public access and attributed the exposure to human error. (europesays.com) Draft documents in the exposed cache used the names Mythos and internal codename “Capybara,” claimed substantially higher scores on coding and reasoning benchmarks versus Opus models, and warned the system could elevate cybersecurity risks by making certain offensive capabilities easier to automate. (the-decoder.com) Developers have reported abnormally fast token burn and rapid rate‑limit drains in Claude Code on forums and GitHub, with some suggesting efficiency regressions or client‑side issues that compound the company’s capacity controls. (macrumors.com)
