At 2 pm German time the working day starts on the US West Coast, and with it the window Anthropic itself treats as peak hours: weekdays from 5 to 11 am Pacific Time, which is 2 to 8 pm in Germany. For that window there is evidence that limits run out faster and that more incidents are reported. There is no evidence that the model reasons worse during it, and Anthropic explicitly denies that. If you schedule demanding work for the morning and know a few settings, you will notice far less of the afternoon.
The impression is common, and I know it from my own work with Claude Code: mornings go smoothly, afternoons bring waiting, aborted runs and answers that miss the point. I wanted to know how much of that can be shown.
What happens at 2 pm
The arithmetic is simple. 5 am on the US West Coast is 2 pm in Germany, in summer and in winter, because both sides change their clocks. Only in the weeks when the US is on daylight saving time and Europe is not does the window move to 1 to 7 pm. That is two to three weeks in spring and one in autumn.
At 2 pm the US East Coast has been at work since 8 am local time, the West Coast is starting, and Europe is still working. For a few hours the working days of two continents overlap. During the German morning it is night in the US.
What is on record
Peak hours are official. At the end of March 2026 Thariq, a member of Anthropic's team, explained in a thread on X that the budget of a five-hour session is used up faster on weekdays between 5 and 11 am Pacific Time. Weekly limits stayed the same. The Free, Pro and Max plans were affected, and by his account about 7 percent of users would hit limits they would not have hit before. That is how PiunikaWeb sums it up; I could not retrieve the thread itself.
For Claude Code the rule has been lifted again. On 6 May 2026, together with new compute capacity, Anthropic announced it was removing the peak hours limit reduction on Claude Code for Pro and Max accounts. For the chat in the browser and the apps I found no announcement that it ends.
The infrastructure is stretched at peak hours. In April 2026 Anthropic told Fortune that demand had grown at an unprecedented rate and that its infrastructure had been stretched to meet it, "particularly at peak hours".
Incidents cluster in the German afternoon. I counted this myself. On 17 November 2026 the Claude status page returned its 50 most recent incidents through its API. They run from 17 August to 16 November 2026, and 42 of them fell on weekdays. In German time those 42 are spread like this:
| Incident reported (German time, weekdays) | Incidents | Share |
|---|---|---|
| midnight to 8 am | 6 | 14% |
| 8 am to 2 pm | 5 | 12% |
| 2 pm to 8 pm | 22 | 52% |
| 8 pm to midnight | 9 | 21% |
The six hours from 2 pm are a quarter of the day and carry a good half of the incidents. The morning, also six hours, had five. Two caveats belong here. An incident on the status page is a disruption, meaning errors, outages or long response times, and no measure of how good an answer is. And I suspect that incidents are more likely to be reported when someone at Anthropic is at their desk, which would favour the US working day further. I cannot prove that.
What Anthropic denies
In September 2025 Anthropic wrote in a postmortem of three issues: "We never reduce model quality due to demand, time of day, or server load." By that statement the model is not throttled when things get busy.
The same report shows that answer quality can still vary for weeks. In August 2025 requests to Sonnet 4 were routed to servers set up for a different context window, 16 percent of them in the worst hour. Two further bugs corrupted output in the same period. None of the three depended on the time of day.
A second report followed in April 2026, this time about Claude Code. On 4 March Anthropic had lowered the default reasoning effort from high to medium and reverted that on 7 April. From 26 March to 10 April a bug hit sessions that had been idle for more than an hour: after that, Claude Code discarded the older thinking on every turn and kept only the most recent block. A system prompt change meant to make answers shorter hurt coding quality from 16 to 20 April. Nothing in it mentions load or time of day either.
My conclusion: when Claude answered worse than usual, the cause was twice a documented change or bug at Anthropic that acted all day. In the afternoon, waiting times, errors and tighter limits come on top, and together that feels like a dumber model.
Two explanations on our side
Both are my own assessment and not measured.
By 2 pm a session that began in the morning is long. It holds discarded approaches, old error messages and files that have long since changed. A model that reads all of that along answers less precisely than in a fresh conversation, at any time of day.
We are also more tired in the afternoon. Prompts get shorter and vaguer, and patience for a second round drops.
What helps
- Put the hard work in the morning. Start architecture, large refactorings and long agent runs between 8 am and 2 pm German time. Keep tasks for the afternoon where an aborted run costs little.
- Status page first, prompt second. When runs abort and answers drag, status.claude.com shows within seconds whether an incident is the reason. That saves the half hour spent rewriting your own prompt.
- Check the reasoning effort. In Claude Code,
/effortshows and changes the level. According to the documentation, Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5 start atmedium. For a single hard question the wordultrathinkin the prompt is enough. - Know your fallback model. If you run Claude Code with
--fallback-modelor thefallbackModelsetting, an overloaded primary model means that this one turn is answered by another, usually smaller model. Claude Code shows a notice when it switches. It is the only case I know of in which a weaker model really answers under high load, and you switch it on yourself. - Start fresh after lunch. A new session with a short summary of where things stand beats continuing the one from the morning.
- Plan for limits. In the chat on Free, Pro or Max, the five-hour budget drains faster between 2 and 8 pm German time. Long conversations with large attachments belong in the morning or the late evening. In Claude Code on Pro or Max this has not applied since May 2026.
- Test the impression. Three or four tasks of your own with a known result, asked once at 10 am and once at 4 pm, show more than a gut feeling. If the answers are equally good and only take longer, that points to load as the cause and against a weaker model.
The post on AI coding one year later explains why I only accept results that have tests behind them anyway. In the afternoon that counts twice.
FAQ
Is Claude really dumber in the afternoon? There is no evidence for that. Anthropic states that it never reduces model quality because of demand, time of day or server load. What is on record for the hours from 2 pm German time is more incidents and, in the chat, limits that are used up faster.
Why 2 pm of all times? Anthropic treats weekdays from 5 to 11 am Pacific Time as peak hours. That is 2 to 8 pm in Germany. In a few weeks in March and at the end of October, when daylight saving time does not match on both sides, it is 1 to 7 pm.
Do peak hours apply to Claude Code too? For Claude Code on the Pro and Max plans, Anthropic removed the peak hours limit reduction on 6 May 2026. Incidents and longer response times still affect Claude Code.
When is the best time to work with Claude in Germany? On weekdays between 8 am and 2 pm. Of the 42 incidents the status page reported on weekdays between mid-August and mid-November 2026, 5 began in that window and 22 between 2 and 8 pm.
Has Anthropic ever made the quality worse? Unintentionally, yes. In August 2025 it was three infrastructure bugs, in March and April 2026 three changes to Claude Code, among them a lowered default reasoning effort. Both times Anthropic published a report and reverted the changes.
How do I tell whether an incident is running right now? From the status page at status.claude.com. It lists current and past incidents with the time and the affected services.
Sources
- Anthropic: A postmortem of three recent issues, three infrastructure bugs in August and September 2025, statement on model quality (17 September 2025)
- Anthropic: An update on recent Claude Code quality reports, three changes to Claude Code with dates (23 April 2026)
- Anthropic: Higher usage limits for Claude and a compute deal with SpaceX, end of the peak hours limit reduction for Claude Code on Pro and Max (6 May 2026)
- PiunikaWeb: Anthropic finally explains why Claude usage limits feel tighter for some users, summary of the thread on peak hours, times, plans, 7 percent (27 March 2026)
- Fortune: Anthropic explains Claude Code's recent performance decline after weeks of user backlash, statement on demand and infrastructure (24 April 2026)
- Claude Status: Incident History, the 50 incidents from 17 August to 16 November 2026, my own count in German time on 17 November 2026
- Claude Code Docs: Model configuration, effort levels and defaults, fallback models under overload