Is Claude down, or am I out of tokens?

3 August 2026 · Written by the developer of MeterTab

Claude Code stops working. From where you are sitting, an Anthropic incident and your own exhausted window look almost identical: things fail, and you do not know whether to wait, switch models, or go and do something else. They call for opposite responses, so it is worth being able to tell them apart quickly.

Check 1: read the error text properly

The two failures do say different things, and it is easy to skim past the distinction when you are annoyed.

	Your usage limit	Their outage
Wording	Names a limit and a reset time	Talks about overload, capacity, or a server error
Shape	Rate limited - about you	Overloaded / 5xx - about them
Retrying helps?	No. It fails identically and costs you nothing but time	Sometimes, after a backoff
Switching model helps?	Sometimes - limits differ per model family	Sometimes - a component can be degraded on its own
Right response	Stop, or continue on credits at API rates	Back off, go get coffee, do not burn your window

If the message hands you a reset time, that is your limit, and no amount of retrying will change it.

Check 2: do you actually have headroom?

Two authoritative places, both take seconds:

/status inside Claude Code - your remaining allocation, right there.
Claude → Settings → Usage in a browser - progress bars for your 5-hour session and weekly limits, with their reset times.

If those show plenty left and things are still failing, it is not you.

Check 3: ask Anthropic's status page

Anthropic runs a public status page at status.claude.com, and it publishes a small JSON summary that is much faster to read than the page:

curl -s https://status.claude.com/api/v2/summary.json \
  | python3 -c 'import json,sys; d=json.load(sys.stdin); \
print(d["status"]["description"]); \
print([c["name"]+": "+c["status"] for c in d["components"]])'

That prints the overall indicator plus the state of each component - and the components are listed separately, so you can see that Claude Code is degraded while claude.ai and the API are fine, or the reverse. Worth knowing before you conclude your setup is broken.

An important negative result: "All Systems Operational" does not prove you are not affected. Status pages lag reality, sometimes by a lot, and per-account or per-region problems may never appear on one at all. Treat green as weak evidence, and red as strong evidence.

The expensive version of this mistake

The reason this matters more now than it used to is that you are frequently not the one hitting retry.

When a long agent run starts failing, the agent's instinct is to try again. During an outage that is a loop that burns real tokens against your window on calls that were never going to succeed - so you can come back to find that a service problem has consumed the budget you were saving for the actual work. It is a genuinely bad interaction: their downtime, your quota.

Which is why the useful move is not just being able to tell the difference yourself, but letting the agent tell the difference. MeterTab exposes both facts over MCP: get_budget for how much of your window is left, and get_service_status for whether Claude itself is degraded. An agent that can call both stops retrying into an outage and waits, instead of grinding your weekly cap into a wall.

The same app watches the status page for you and can put an icon in the menu bar only when something is actually wrong - so "why is this failing?" has an answer before you start debugging your own machine.

Quick reference

Error names a reset time → your limit. Wait, or enable usage credits.
Error says overloaded or 5xx, and /status shows headroom → their problem. Back off.
Status page red on the Claude Code component → confirmed their problem. Stop retrying.
Status page green but only you are affected → check your network, proxy, VPN and any corporate TLS interception before assuming an outage.
Either way, stop the agent from retrying blindly - a PreToolUse hook can enforce that.

MeterTab shows your remaining budget and Claude's service status in one menu-bar glance, and lets your agents ask about both.

Download on the App Store

Status endpoint verified live on 3 August 2026. Anthropic can change these URLs; the status page itself is the durable link.