Claude token limit per day: the ceilings that actually exist

There is no daily limit

On a Claude subscription, usage is measured against a rolling five-hour session limit and weekly limits. On the Claude API, throughput is measured per minute and spend is capped per calendar month. Neither system has a day-shaped bucket anywhere in it, which is why no official page states one and why every number you can find for it is somebody working backwards from their own bill.

Every ceiling that does exist

• API monthly spend caps: $500 on Start, $1,000 on Build, $200,000 on Scale. Custom tier has none. You can set your own limit below your tier cap on the Billing page.
• Per-minute limits are per model, so several models can run up to their respective limits at once.
• Capacity is a token bucket: it replenishes continuously rather than resetting at an interval, which is why bursts fail even under the stated number.

The context window, which is what most people mean

When someone says they hit a Claude token limit, they usually mean the conversation got too big rather than the account got too used. This is the per-request ceiling: how much the model can hold at once, re-sent in full on every turn.

• On claude.ai, paid plans get up to a 1M token context window on the newest models, with others at 500K or 200K.
• On the Message Batches API, Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5, and Sonnet 4.6 support up to 300k output tokens using the output-300k-2026-03-24 beta header.
• Running a 1M-context model through a provider that serves it at 200K, such as some cloud platform deployments, gives you the smaller window and the earlier compaction that goes with it.

# the Models API returns max_input_tokens, max_tokens and capabilities per model
curl -sS https://api.anthropic.com/v1/models \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01"

Where the window actually gets cut

In Claude Code you rarely meet the model ceiling directly, because auto-compaction summarises older history first. With nothing configured, it compacts when the conversation reaches the model context limit, with documented exceptions: cloud sessions compact as they approach it, Sonnet 4.6 and Opus 4.6 without extended context compact at the 200K boundary, and so do Opus 4.8 and Opus 5 when they run with a 200K window.

/autocompact 500k     # this session and later ones
/autocompact auto     # back to the window tuned for your model
/context              # what is filling it right now, by category
/compact focus on the auth refactor and the decisions we made

Building a daily budget yourself

If you want a daily number, you have to impose one. The mechanisms differ by how you are billed, and only one of them is enforced rather than advisory.

For a genuine per-day figure across everything you run, the practical answer is to measure locally rather than wait for a bill. Every agent writes a session transcript with token counts per turn, and daily rollups over those are exactly what ccusage daily produces.

npx ccusage@latest daily --since 20260801
npx ccusage@latest daily --breakdown --json | jq '.[-1]'

Which ceiling you actually hit

Questions people ask

Keep reading