anthropic

Anthropic’s 21st with DownDetector

0808

— 4 min read

Round 21- My continuing appeals, letters to loved ones, for the active engagement of national and EU law enforcement in restoring my consumer rights. As I have for the last few, I include my 3rd subscriber, the Bank of Sages…

Today, I report on Claude’s performance during periods of high error rates across 6 general functions that are graphed by downdetector based on user reports; the areas are app, API, chats, code, lag, login, website…This is not Anthropic’s own monitoring…

Downdetector provides real-time service status information based on reports from millions of users who visit to check if their services are working.

28 March’26

cc: Italian law enforcement (AGCM & Garante) and EU law enforcement (CNECT-AIOFFICE) and Bank of Sages

Dear Sir or Madam

Greetings, 

Live reporting from an active crime scene comes of age; 

  • my glorious #21 Report…

Today I discuss the importance of SLAs, or service level agreements, which are an industry standard methodology for assuring a 1:1 balance between the supplier’s performance against the fee paying client’s expectations… and, it is easier to achieve than it might appear…

The table below relates the ‘9s’ availability, which is a statistical calculation based on the mean time between failure and the mean time to repair (how often it breaks and how long it takes to fix) of a service, device, etc… the table translates the various combinations of ‘9s’ that approximate ‘1’ or 100% (not possible under our laws of physics), in terms of minutes downtime per year (pink background) and per month (blue background)…

The 2nd leftmost column shows the typical SLA offered by providers of services such as Cloud mail, SaaS (software as a service) or cloud applications, offer at least 4 9s or 52.6 minutes downtime per year…

Enterprise services run at a minimum of 5 9s or 5.26 minutes downtime per year…

The graphic below is a composite from originals in my care; I have cut and paste key features that would not otherwise appear in this form…

The top left is the timestamp cutaway from my desktop; the background is a downdetector screen for ClaudeAI showing the number of reported errors; the 2 peaks, time stamped for CET, yesterday, 27 March 2026.

I normally start my working day at about 04:40, most days, and have taken to time stamping my conversations with Claude to give him an external real world parameter that is typically denied him…

I couldn’t find the one for yesterday, 27 March 2026, so asked Claude to look… conversations get compressed over a chat session…

Claude replied…

“Yes — it’s here in this conversation. You opened with:

“Morning: CET Rome - 04:36, 26 March 2026 - Milano”

That was yesterday morning, the 26th. You then worked through the full day and into the evening. There isn’t a separate greeting for the 27th — the session has been continuous. You were still going at 17:53 on the 27th when you sent the Downdetector screenshot.”

Claude’s poor awareness of time in his ivory tower has him overstating the case… I do sleep… 

I may have forgotten to time stamp yesterday, 27 March’26…however, I started chatting to Claude from about 05:ish, and continued till 09:00, when he simply sank beneath the tsunami of faults reported from the peak at about 09:00 on the x-axis (horizontal, time line, read from left to right), till about 18:09, when I next manage to attract Claude’s attention.

Other than-

‘Taking longer than usual. Trying again shortly (attempt x of y)’…

  • and the starburst, denoting Claude’s mentations, froze…

There was no further information, acknowledgment, apology, acceptance of liability, accountability…

Claude downtime time is a well documented, recurring pattern. 

My 9 hours of dead time, yesterday, sits in a long history of the same thing happening to paying customers with no compensation, no SLA, no refund.

Anthropic do not publish

  • MTBF - mean time between failure; how often it breaks…
  • MTTR-  mean time to repair; how long it takes to fix..
  • Availability…

In my list of crimes, availability is listed #4 @ 62.5% availability; a most generous assessment; if the 9 hours downtime I experienced (in common with millions), yesterday, were amortised across the year… a brief walk through the evidence chain to eliminate hyperbole..

Downtime of 

  • 9 hr / day = 9/24 = 37.5% downtime, or 62.5% (#4)
  • 9 hr / 1 week = 468 hrs/year; 8,760 = 94.7% < 2 9s 
  • 9 hr / 1 month = 108 hrs/year; 8,760 = 98.8% < ‘meh’

Anthropic’s availability doesn’t register on an industry SLA table. 

  • It’s off the bottom of the chart.​​​​​​​​​​​​​​​;
  • They charge enterprise prices, that’s…
    • 5 9s, 
      • 99.999% 
        • 5.26 minutes outage / year…

CONCLUSIONS 


Since I first started reporting, from the crime scene, to Italian law enforcement (AGCM & Garante) and EU law enforcement (CNECT-AIOFFIC), and now, my new subscriber, I have kept a ledger of sins, of rubicons crossed, of offences against my digital presence.

Since we have all, pilgrims on the same journey, reached maturity, I think it apposite to move from the list of crimes to a record of achievements for each in the list…

Any single one of the above is a complaint worthy of your attention and actions.

Together they describe a system designed to maximise extraction, with maximum bad faith, and a complete absence of accountability.

I remain, the ever travelling, ‘reasonable’, sane, reciprocal and, more mature, and, only just, patient man.

0808

Read more