How many AI coding agents can I run in parallel on one API key?
One Standard Compute key accepts 300 to 2,000 requests a minute on personal plans, enough for about 25 to 160 agents at 10 requests a minute each, but the monthly compute budget runs out long before that: one agent on Claude Sonnet 5.5 costs about $3.53 on a median day at list prices. On Anthropic and OpenAI keys, every key in the organization shares one rate limit, and on Claude Pro or Max every session shares one 5-hour and one weekly limit.
Requests per key, by plan
Each key has two counters: requests in the last minute and requests in the last 10 minutes. There is no 5-hour or weekly window, no tokens-per-minute limit, and on paid plans no limit on how many requests run at once.
| Plan | Per min | Per 10 min | Agents at 10/min |
|---|---|---|---|
| Starter $19 | 300 | 2,500 | 25 |
| Economy $39 | 600 | 5,000 | 50 |
| Standard $89 | 1,000 | 8,000 | 80 |
| Pro $249 | 2,000 | 16,000 | 160 |
| Pro Plus $499 | 2,000 | 16,000 | 160 |
| Growth $999 | 3,000 | 24,000 | 240 |
| Scale $2,499 | 4,000 | 32,000 | 320 |
Per API key, from the gateway's plan settings on October 14, 2026. Over 10 minutes Starter allows 2,500 requests, so its sustained rate is 250 a minute, and the last column divides that by 10 requests a minute per agent. Pro Plus, Growth and Scale are company plans.
What one agent actually sends
- A median coding-agent session on our gateway is 16 requests in 8.1 minutes, about 2 a minute. Every round of tool results is a new request that carries the whole conversation again.
- Anthropic's own sizing guide for Claude Code plans 5 to 7 requests a minute per developer for teams of 1 to 5, at the organization level.
- Subagents multiply this. Claude Code runs up to 20 subagents at once in one session by default, and each one sends its own requests. Codex sets the number with
agents.max_concurrent_threads_per_session. - So the table above assumes 10 requests a minute for an agent working without pause, five times the median session's pace.
Vendor limits when the agents use your own key
| Where the agents run | Limit | Shared by | Source |
|---|---|---|---|
| Anthropic API, Start tier | 1,000 requests and 2M uncached input tokens a minute on Sonnet 5.5 or Opus 5.5; $500 monthly spend cap | Every key in the organization | Anthropic |
| Anthropic API, Build / Scale | 5,000 / 10,000 requests and 5M / 10M input tokens a minute; $1,000 / $200,000 cap | Every key in the organization | Anthropic |
| OpenAI API, Build / Launch / Grow | 5,000 / 10,000 / 15,000 requests and 1M / 4M / 40M tokens a minute on the Astra, Sol and Terra models | The organization and project | OpenAI |
| Claude Pro or Max | A session limit that resets every five hours, plus a weekly limit | claude.ai, Claude Code and Claude Desktop together | Claude Help Center |
| ChatGPT plans in Codex | A five-hour and a weekly limit on Plus; weekly only on Pro | Codex CLI, IDE, app, web and cloud tasks | Codex usage limits |
| OpenCode Go / Go Plus | Per model: 20% of the monthly dollar limit per five hours, 50% per week | One single-member workspace | OpenCode |
| Standard Compute | 300 to 4,000 requests a minute by plan; monthly compute budget | Requests: each key. Budget: the plan | This page |
Checked on each vendor's own page on October 14, 2026. Anthropic says new organizations may start on a lower Evaluation tier, that 60 requests a minute can be enforced as one a second, and that cache reads don't count toward input tokens on most models. OpenAI counts unsuccessful requests toward the per-minute limit. Tier tables for every provider: LLM API rate limits.
What happens when parallel agents hit a limit
- Request limit: HTTP 429
rate_limit_exceededwith a Retry-After header of 60 or 600 seconds, depending on which window filled. The message names the request-rate limit, so you can tell it from a budget stop. - Retries count. A rejected request still adds to the window, so an agent that retries without waiting keeps the window full.
- Budget: at 100% of the monthly budget, new requests get HTTP 402
budget_exhaustedand requests already running finish. OpenAI SDKs don't retry a 402, so agents stop with a clear error. - Upstream 429s from a model provider are retried by the gateway and can move to another provider of the same model, so one provider's limit doesn't stop your agents.
Split keys and budgets per agent
A company workspace is free to create and lets you give each agent, worktree or fleet its own key, so you can see and cap what each one spends:
- Every key has its own request counters, so a second key doubles the requests you can send.
- Personal keys can spend from their owner's plan: one monthly budget, several keys. Use this for your own parallel sessions.
- Application keys belong to a team or the company and pay for themselves: their own plan, or the company pay-as-you-go balance with a monthly spend limit per key. Use this for an always-on fleet that must not drain your coding budget.
- When one key's plan runs out, keys on other plans keep working. Usage and the request log show each key's spend and the model that served each request.
What N agents cost for a day
| Agents | Sonnet median | Sonnet heavy | Luna median | Budget for 22 days |
|---|---|---|---|---|
| 1 agent | $3.53 | $28.97 | $0.23 | Standard ($95) |
| 3 agents | $10.59 | $86.91 | $0.69 | Pro ($269) |
| 5 agents | $17.65 | $144.85 | $1.15 | Pro Plus, company ($549) |
| 10 agents | $35.30 | $289.70 | $2.30 | Growth, company ($1,119) |
Day costs for Claude Sonnet 5.5 and GPT-6 Luna at list prices from our gateway's traffic data (Agent Burn Index, September 13 to October 12, 2026): the median and the heavy (90th-percentile) coding-agent day. The index measures one account's day, which this table treats as one agent's day, so read it as an estimate. Last column: the smallest Standard Compute plan whose compute budget covers 22 median Sonnet 5.5 days for every agent if every request goes to Sonnet 5.5; with routing, cheaper models take part of the work.
The catch
The request limit is high on purpose, so the budget is what stops parallel agents. On the Standard plan ($89 a month, $95 compute budget), five agents sending every request to Sonnet 5.5 use it up in about 5 median days. Many agents on a flat plan work when simple steps go to cheaper models, or when the budget is sized for the whole fleet.
Where Standard Compute fits
Standard Compute is one OpenAI- and Anthropic-compatible API at https://api.stdcmpt.com/v1 for Claude Code, Codex CLI, OpenCode, Cline, OpenClaw and Hermes, with Claude, GPT, Gemini, DeepSeek, GLM and Kimi behind one key. Plans run from $19 to $249 a month for individuals and up to $2,499 for companies, each with a fixed compute budget and nothing billed on top. With the model set to standardcompute, routing sends simple agent steps to cheaper models, so a parallel fleet's budget lasts longer.
FAQ
Do more API keys give me more requests per minute?
At Anthropic and OpenAI, no: rate limits apply to the organization (and at OpenAI also the project), so a new key shares the same pool. On Standard Compute, yes: each key has its own per-minute and per-10-minute counters. The budget is separate: keys that share one plan share its monthly budget.
Can I run Claude Code, OpenCode and Codex at the same time on one key?
Yes. They all count toward the same two request counters and the same monthly budget. OpenCode and Codex CLI use https://api.stdcmpt.com/v1; Claude Code uses https://api.stdcmpt.com without /v1, because it adds /v1/messages itself. Give each tool its own key if you want to see what each one spends.
Is there a limit on how many requests run at once?
Not on paid plans: only the per-minute and per-10-minute counts apply, so long streaming requests in parallel are fine. Free usage limits how many requests run at once and answers HTTP 429 free_concurrency_limit when that is full.
How many OpenClaw or Hermes agents can share one key?
Always-on assistants sit idle between tasks most of the day, so their request rate is low and the ceiling rarely matters. Their cost is the long context sent with every turn. Put a fleet on its own application key with its own plan or spend limit, so it can't use up the budget your coding agents need.
I run agents in git worktrees with Conductor or Claude Squad. What changes?
Each worktree runs an ordinary agent session, so each one sends its own requests. Point every session at the same key to share one budget, or set a different key per worktree to split spend by task.
What's the cheapest way to run five agents all day?
Mix models. A median agent day costs about 15 times less on GPT-6 Luna than on Sonnet 5.5 at list prices. Keep a frontier model for the main agent, run subagents on a cheaper model, and pick a plan whose budget covers your working days.