GuidesRunning & maintaining

OpenRouter "Key limit exceeded (daily limit)" with credit left

A 403 from OpenRouter when a key's own cap or a workspace budget is spent, not your balance. How to see which one, when it resets, and how to raise it.

November 17, 2026The Everpod team
The short answer

“Key limit exceeded (daily limit)” is a 403 from OpenRouter saying the spending cap set on that one API key is used up, not that your account is out of credit. “Budget limit exceeded (monthly limit)” and its variants are the same refusal from a budget set on a workspace or an organisation member. Adding credits fixes neither. A daily, weekly or monthly cap clears by itself at midnight UTC at the start of the next period; otherwise raise the key’s limit on its page in OpenRouter’s key settings, or, for a workspace or member budget, an organisation admin raises it. GET /api/v1/key shows which budget is spent and when it resets.

The wordings, and what agents call them

OpenRouter refuses these with HTTP 403 before the request reaches any model provider. The message names the budget and, in brackets, its period. Forms seen as of November 2026:

Agents have often shown something else. Hermes Desktop put up “Out of credits — OpenRouter” and an “Add credits” button for the daily key cap, an issue still open in November 2026. Earlier, Hermes and OpenClaw both read the organisation budget message as a failed login and told the user to check the key; both were fixed, Hermes in September 2026 and OpenClaw on its main branch in May. So an agent saying you are out of credit, or that your key is bad, while OpenRouter’s credits page shows a balance and the key works elsewhere, is worth checking against the message itself in the agent’s log.

Three budgets, checked in order

OpenRouter’s limits reference checks every request against up to three kinds of cap, in this order, and the first one that is spent blocks it:

  1. Guardrail budgets. A dollar cap with a daily, weekly or monthly reset, set by an organisation admin on a workspace’s guardrails, for every member, for one member, or for one key. A member’s budget counts everything spent through all of that member’s keys.
  2. Workspace budgets. A cap on the whole workspace’s spend, daily, weekly, monthly or lifetime. A lifetime budget never resets.
  3. The key’s own limit. The credit limit you choose when creating a key, resetting daily, weekly, monthly or never. Keys that OpenRouter creates for you during onboarding come with a $100 limit, so a key you never set a limit on may still have one.

Your account’s credit balance sits under all three. Running it down gives a different error, the 402 covered in OpenRouter’s “requires more credits” message, so a 403 with one of the wordings above means the balance was not what stopped the request.

Finding out which one is spent

The documented 403 body carries the answer in error.metadata: limit_source is openrouter_key_limit, openrouter_workspace_budget or openrouter_guardrail_budget, and beside it are the scope that blocked, limit_usd, spend_usd and resets_at, a UTC time, or null for a budget that never resets. Many agents log only the message, so the quicker check is to ask OpenRouter about the key directly, from the machine the agent runs on and with the variable the agent reads:

curl https://openrouter.ai/api/v1/key \
  -H "Authorization: Bearer $OPENROUTER_API_KEY"
Copied

The response does not print the key; its label is shortened. In its data, as named in the endpoint’s reference:

A 401 from the same command instead means the agent’s key is not one OpenRouter accepts, a separate problem with its own wordings, in OpenRouter’s 401 errors.

Waiting for the reset, or raising the cap

Every periodic cap resets at 00:00 UTC: daily ones each midnight, weekly ones on Monday, monthly ones on the first of the month. On a server or a laptop outside UTC that is not your midnight; resets_at gives the exact time. Until then, OpenRouter’s docs say retrying fails the same way, and the 403 carries no Retry-After header.

“Contact your org admin” in the message means exactly that: the cap was set by whoever runs the organisation’s OpenRouter account, and it is theirs to raise or keep.

When the key still has some limit left

A key limit can stop a request before limit_remaining reaches zero. OpenRouter prices a request at the most it could cost before running it, and a key with less left than that gets the 402 “This request requires more credits, or fewer max_tokens” with a pointer to a higher key limit; the arithmetic is in the requires-more-credits guide. One OpenCode user met “Key limit exceeded (total limit)” only on DeepSeek V4 Pro, with credit on the account and limit left on the key, while smaller models worked through the same key. The cause was not settled in the report, but the pattern to check is the same: the larger model’s worst-case cost against what the key has left.

What an agent on a server does meanwhile

Every request on that key is refused, whatever the model: the cap belongs to the key, the member or the workspace, and the request is refused before it reaches a provider. A fallback to another OpenRouter model on the same key fails the same way; a fallback that does not go through that key is the kind that keeps the agent answering. Background work inside the agent goes with it; one Hermes report found the summarising that keeps a long session within the model’s context also stopped, leaving the session unusable.

An agent that treats the 403 as a passing failure can make it worse: in the OpenClaw report, it retried the same model “thousands of times” and appeared to hang. Nothing goes through before the reset, so a scheduled job or an agent left running overnight on a daily cap simply fails until midnight UTC. Before leaving one unattended, compare the key’s usage_daily on a normal day with its limit and its reset, and set the cap to what the agent really spends, so it stops for a runaway rather than for an ordinary day. If the message reads “Provider returned error” instead of one of the wordings above, the refusal came from the model’s own provider, not from a budget.

Run Claude Code, Codex, OpenCode, Pi, Hermes or OpenClaw on an always-on developer pod.

A developer pod is a permanent cloud computer (VPS), set up and secured for you, with your pick of Claude Code, Codex, OpenCode, Pi, Hermes and OpenClaw installed. Run more agents at once than your laptop can hold, and keep them working when it’s closed. From $24 a month, built in about ten minutes.