Free tool · no signup · nothing leaves this page

How many agents can your org run before the 429s start?

Rate limits are not one number, they are three meters running at once, and only one of them ever binds. Pick your Anthropic tier, your OpenAI tier, or your Claude plan, describe one agent's workload, and this computes the org-wide concurrent-agent ceiling from published limits, names the meter that fires first, and prices the next rung against more seats and against spreading the fleet across keys.

01 · The planner

Your plan, and one agent

Provider
Model class

Concurrent, not sequential. Two agents running for an hour is two agents, and this is the number people underestimate once agent teams are in play.

One agent's workload

Concurrent-agent ceiling
0 agents

fits

0agents wanted
0agents allowed
0%headroom left
-binding meter
Every meter, and how close you are to it.checked aug 2026
MeterPublishedYour demand AllowsUsed

The ladder, against the same workload.

The highlighted row is the cheapest rung that covers what you are asking for.

Assumptions, and which numbers are derived

An API ceiling is arithmetic. Divide each published meter by what one agent draws from it per minute, and the smallest quotient is how many agents that tier carries. A subscription plan has no published per-minute meter at all, so its ceiling comes from the published weekly allowance spread across a forty hour week, and from the rolling five-hour window.

  • Published. Anthropic's Claude API limits, read from platform.claude.com/docs/en/api/rate-limits on 20 August 2026. The tiers are Start, Build, Scale and Custom, and the Tier 1 to Tier 4 numbering people still search for no longer exists.
  • Published. Anthropic excludes cache_read_input_tokens from ITPM on every current model, so only uncached input and cache writes count against the input meter. Anthropic's own worked example: a 2M ITPM limit at an 80 percent cache hit rate effectively processes 10M input tokens a minute. That is why the cache slider moves this ceiling further than the tier does.
  • Published. Claude Fable 5 is metered on its own bucket and far lower than the rest: 500k ITPM and 100k OTPM on Start, against 2M and 400k for Opus 5, Sonnet 5 and Haiku 4.5, which share one bucket. Fable also has its own RPM on Build and Scale.
  • Published. OpenAI's tier limits, read from developers.openai.com/api/docs/models/gpt-5.6 and the GPT-5.2 model page on 20 August 2026, with the tier qualification thresholds from the rate limits guide. Both flagship models carry identical numbers, which is why this ladder has no model selector.
  • Published. The Claude subscription weekly floors: roughly 40 Sonnet-hours a week on Pro, 140 Sonnet plus 15 Opus on Max 5x, and 240 Sonnet plus 24 Opus on Max 20x. Anthropic published these as ranges (Pro 40 to 80, Max 5x 140 to 280, Max 20x 240 to 480); this tool uses the low end, because a ceiling you plan against should be the one that holds in a bad week.
  • Published. On Max, Team premium and seat-based Enterprise premium seats, Claude Fable 5 may draw at most 50 percent of the weekly pool. It is a slice of the same budget, never a second allowance, and at roughly five times a Sonnet hour it spends that slice quickly.
  • Assumption, not a published fact. OpenAI's TPM is treated here as a single combined meter counting input, cached input and output alike. OpenAI documents TPM as a token meter without publishing a cache exclusion, so this is the conservative reading. If your cached input is in fact excluded, your real ceiling is higher than the number above.
  • Derived. The default agent workload: 2.74M tokens per agent-hour, or about 45k tokens a minute, of which 2.48M an hour are cache reads, plus 32k an hour of output. It comes from a worked twenty-turn agent session taken at two sessions an hour, and it is the same figure the Claude Code plan picker runs on. Request rate is set at 2 a minute on the assumption of roughly three API calls per agent turn.
  • Derived. The five-hour window capacity: Anthropic publishes a multiplier and no number, so Pro is set at 2.5 Sonnet-equivalent hours per rolling window and scaled by the published multipliers to 12.5 and 50. Pro's figure is calibrated against the widely reported experience that Pro carries one to three hours of Claude Code a day.
  • Derived. Pro's weekly Opus allowance. Anthropic published no Opus figure for Pro, so it is scaled from Max 5x's 15 hours by the ratio of the two Sonnet floors, giving about 4 hours a week.
  • Derived. Model burn multipliers of 2.5x for Opus-class and 5x for Fable against a Sonnet hour, taken from the list price ratios, which are the closest public proxy Anthropic gives for relative allowance burn.
  • Simplification. A forty hour working week with every agent running continuously. Real agents idle between turns, so a plan ceiling computed this way is pessimistic; an API ceiling computed this way is not, because a per-minute meter cares about your peak minute rather than your average one.
  • Not modelled. Anthropic's acceleration limits, which fire below the published ceiling when usage climbs sharply; the shorter enforcement intervals that turn a per-minute limit into a per-second one; workspace-level sub-limits; batch API limits, which are a separate pool; Priority Tier; and the fact that Anthropic estimates input tokens at the start of a request and corrects during it.

All published limits checked August 2026. Both vendors move these, sometimes upward without notice; read your own console before you plan a quarter around a number on this page.

02 · Why one meter always binds

Three limits, and you only ever meet one.

Every capacity conversation starts with the wrong number. The tier you are on is not your ceiling; the smallest of its meters divided by your workload is.

The meters are not interchangeable

An Anthropic API tier publishes three limits at once: requests per minute, input tokens per minute, and output tokens per minute. They are enforced independently, which means your real ceiling is set by whichever one your particular workload exhausts first, and for a coding agent that is almost never the request meter. Agents make few calls carrying enormous context, so a fleet that is nowhere near 5,000 requests a minute can be sitting on the input meter. Change the workload and the binding meter moves: an agent doing bulk generation rather than repository reading binds on output instead, at a tenth of the token volume.

OpenAI splits it differently, with a single combined tokens-per-minute meter alongside requests. That sounds simpler and is materially harsher for agent work, because everything counts: the prompt, the context you resend on every turn, and the reply. A 500,000 TPM Tier 1 account running agents at 45,000 tokens a minute each carries roughly ten of them. Ten is not a lot when a single developer runs two.

Caching is the lever, not the tier

Anthropic excludes cache reads from the input meter entirely. Cache writes count, uncached input counts, cache reads do not, and for a coding agent working a warm repository the cache reads are the overwhelming majority of the input. Anthropic's own documentation works the example: a two million ITPM limit at an eighty percent cache hit rate effectively processes ten million input tokens a minute. In ceiling terms, moving a fleet from a cold cache to an eighty percent hit rate is worth more than a full tier upgrade, and it costs nothing but structuring your prompts so the stable parts sit above the breakpoint.

This is also why two organizations on the same tier can report completely different experiences with 429s. The one that reuses system prompts, tool definitions and repository context across turns is running five times the fleet of the one that rebuilds its prompt every request, on identical published limits.

A subscription has no per-minute meter at all

Claude Pro, Max 5x and Max 20x are not rate-limited in requests per minute. They are budgeted in a rolling five-hour window and a weekly cap, so the failure mode is different in kind: you do not get a 429 in the middle of a burst, you get a Thursday where the account is simply finished until the reset. Expressed as a sustained concurrent-agent number, Max 5x's published 140 Sonnet-hours a week works out to about three and a half agents running continuously through a forty hour week, and the five-hour window holds it nearer two and a half in any single afternoon. On Opus, divide both by 2.5 and then check the separate Opus allowance, which is the tighter of the two.

Seats scale that linearly, which is the one genuine advantage a subscription fleet has: every account carries its own window and its own weekly pool, so ten seats is ten independent ceilings rather than one larger one. What nothing in the CLI does is move a session between them when one runs out.

A second key is not more capacity

Rate limits are enforced per organization, not per key, so issuing a second key inside the same organization buys nothing at all. Workspace limits are a way to reserve capacity for one team rather than add any, and the organization ceiling still applies even when the workspace limits sum to more. Real extra capacity means a second organization, a second provider, or both, which turns the problem into routing: see LLM failover for the failure semantics and LLM gateway pricing compared for what the routing layer costs.

03 · The published numbers

Everything the planner runs on, written out, so you can check it against your own console rather than trusting a slider.

Anthropic Claude API, per model class.checked 20 aug 2026
TierMonthly spend capModel RPMITPMOTPM
Start$500Opus 5 / Sonnet 5 / Haiku 4.51,0002,000,000400,000
Start$500Fable 51,000500,000100,000
Build$1,000Opus 5 / Sonnet 5 / Haiku 4.55,0005,000,0001,000,000
Build$1,000Fable 52,0001,500,000300,000
Scale$200,000Opus 5 / Sonnet 5 / Haiku 4.510,00010,000,0002,000,000
Scale$200,000Fable 54,0004,000,000800,000

Source: platform.claude.com/docs/en/api/rate-limits, read 20 August 2026. Opus 4.x and Sonnet 4.x carry their own combined buckets not shown here. Cache reads do not count toward ITPM on any of these models. Above Scale is a Custom tier arranged with sales.

OpenAI API, flagship models (GPT-5.6 and GPT-5.2 are identical).checked 20 aug 2026
TierQualifies atMonthly usage limit RPMTPM (combined)
Tier 1$5 paid$100500500,000
Tier 2$50 paid$5005,0001,000,000
Tier 3$100 paid$1,0005,0002,000,000
Tier 4$250 paid$5,00010,0004,000,000
Tier 5$1,000 paid$200,00015,00040,000,000

Sources: developers.openai.com/api/docs/models/gpt-5.6 and the GPT-5.2 model page for the per-tier limits, developers.openai.com/api/docs/guides/rate-limits for the qualification thresholds and monthly usage limits. All read 20 August 2026. Smaller models carry their own, usually higher, limits.

Claude subscription plans. Weekly figures are Anthropic estimates, shown as published ranges.checked aug 2026
PlanPrice / moSonnet hours / wk Opus hours / wkFable 55-hour window
Claude Pro$2040 to 80not publishedusage credits, outside the planbaseline
Claude Max 5x$100140 to 28015 to 35up to 50% of the pool5x Pro
Claude Max 20x$200240 to 48024 to 40up to 50% of the pool20x Pro

The planner uses the low end of each range. There is no published per-minute meter on these plans; the concurrent-agent numbers it shows are the weekly allowance spread across a forty hour week, and a derived five-hour window capacity. Detail in the Claude Code rate limits guide.

04 · Questions

The ones that decide the ceiling.

Longer versions live in Claude Code rate limits and LLM failover.

It depends almost entirely on your cache hit rate, because Anthropic excludes cache reads from the input-tokens-per-minute meter. On Build tier with Claude Sonnet 5 the published limits are 5,000 requests per minute, 5,000,000 input tokens per minute and 1,000,000 output tokens per minute. A coding agent drawing 45,000 total input tokens per minute at an 80 percent cache hit rate presents 9,000 uncached, so the input meter alone allows roughly 555 concurrent agents. Run the same agents with a cold cache and that falls to about 111. The tier did not change; the caching did.

As of August 2026 Anthropic names them Start, Build, Scale and Custom, not Tier 1 to Tier 4. For Claude Opus 5, Sonnet 5 and Haiku 4.5 the published limits are 1,000 RPM with 2M ITPM and 400k OTPM on Start, 5,000 RPM with 5M ITPM and 1M OTPM on Build, and 10,000 RPM with 10M ITPM and 2M OTPM on Scale. Claude Fable 5 is metered separately and much lower: 500k ITPM and 100k OTPM on Start. Each tier also carries a monthly spend cap: $500 on Start, $1,000 on Build, $200,000 on Scale.

OpenAI runs five paid usage tiers that you move up automatically as cumulative spend clears a threshold: $5 for Tier 1, $50 for Tier 2, $100 for Tier 3, $250 for Tier 4 and $1,000 for Tier 5. For the flagship GPT-5.6 and GPT-5.2 models the published limits are 500 RPM and 500,000 TPM on Tier 1, 5,000 RPM and 1M TPM on Tier 2, 5,000 RPM and 2M TPM on Tier 3, 10,000 RPM and 4M TPM on Tier 4, and 15,000 RPM and 40M TPM on Tier 5. Unlike Anthropic's split input and output meters, TPM is a single combined token meter.

No. Rate limits are enforced at the organization level, not per key, so a second key inside the same organization draws from exactly the same bucket. What does raise the ceiling is a second organization, a second provider, or a workspace-level split that reserves capacity rather than adding it. A gateway in front of several credentials turns that into automatic routing instead of a developer noticing a 429 and switching by hand.

Three reasons, all of them about time resolution. A per-minute limit is usually enforced over a much shorter interval, so 60 requests per minute can behave like one per second and a burst trips it while the minute-level average looks fine. Anthropic also applies acceleration limits when an organization's usage climbs sharply, which fire even below the published ceiling. And input token limits are estimated at the start of a request and corrected during it, so a request that turns out larger than its estimate can push you over after it was admitted.

05 · Headroom you can watch

This planned the ceiling.
Continuum shows you the meter.

Live gauges on every account and every key, side by side, so the wall arrives as a rising bar rather than a 429. Run several accounts and several providers at once, with the routing in the tool instead of in a developer's head.

one command · reads local history · sends nothing on its own