Every AI application that grows eventually meets the same message: HTTP 429, rate limit exceeded. Rate limits decide how many requests and tokens your company may send per minute, and spend caps decide how much it may spend per month. They are rarely in a project plan, yet they decide whether a launch, a batch run or a busy Monday morning works. We read the official 2026 documentation of Anthropic, OpenAI, Google, Mistral and Manus and compared tiers, limits, caps and error handling.
Two kinds of limits: per minute and per month
All providers combine two mechanisms. Rate limits cap traffic over a short window: requests per minute (RPM) and tokens per minute (TPM). Spend limits cap the monthly bill. Both are set per organisation, not per user or API key, so every application and every developer in the same organisation draws from one pool unless you split it with workspaces or projects.
The details differ in ways that matter. Anthropic counts input tokens (ITPM) and output tokens (OTPM) separately; OpenAI and Mistral use one token limit for input and output together. Anthropic uses a token bucket that refills continuously instead of resetting at the start of each minute, and warns that 60 RPM can be enforced as one request per second. Short bursts can therefore fail even when the minute total is under the limit.
Anthropic: Start, Build, Scale and Custom
Anthropic places organisations on a tier automatically, based on usage history and account standing. New organisations may start in an Evaluation tier below the standard numbers while their history builds up. The current Claude 5.5 models share the same standard limits per tier: 1,000 RPM, 2 million ITPM and 400,000 OTPM on Start, 5,000 RPM with 5 and 1 million on Build, and 10,000 RPM with 10 and 2 million on Scale. Claude Fable 5.x has its own, lower limits. The Custom tier has no published ceiling and no monthly cap.
Source: Anthropic, "Rate limits" (platform.claude.com/docs/en/api/rate-limits), read 27 October 2026. Messages API standard limits; Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 have identical limits per tier, each with its own bucket. RPM: 1,000 / 5,000 / 10,000 (Fable 5.x: 1,000 / 2,000 / 4,000). Values in millions.
Each model row has its own bucket. Opus 5.5 and Sonnet 5.5 do not share a limit, while older models share one (all Opus 4.x versions together, for example). Routing simple requests to Haiku 5.5 therefore adds capacity instead of consuming the Opus budget. The Message Batches API has separate limits as well: 200,000 queued batch requests on Start, 500,000 on Scale, up to 100,000 requests per batch.
Monthly spend caps compared
The monthly cap is the limit most teams discover by accident. Anthropic caps Start at $500, Build at $1,000 and Scale at $200,000 per month. OpenAI moves organisations up automatically as their total credit purchases grow: $5 unlocks Build ($500 a month), $100 Launch ($5,000) and $500 Grow ($200,000). Google ties Gemini tiers to the linked Cloud billing account: Tier 1 needs an active billing account and is capped at $250, Tier 2 needs $100 paid plus three days ($2,000), Tier 3 needs $1,000 paid plus 30 days ($20,000 to $100,000 and more).
Sources: Anthropic "Rate limits", OpenAI "Rate limits" guide (developers.openai.com), Google "Gemini API rate limits" (ai.google.dev), all read 27 October 2026. Anthropic Custom tier has no cap; Gemini Tier 3 is listed as $20,000 to $100,000+. Mistral publishes no tier caps; its limits are shown in the console.
enforced_spend_limit_reached and no retry-after header. SDK retries fail until you move up a tier. A spend limit you set yourself below the cap returns HTTP 400 instead. Check where you stand before a product launch or a large batch job, not during it.
OpenAI: Build, Launch and Grow
OpenAI reduced its usage tiers to three paid levels plus a free tier ($100 a month in allowed regions). Standard limits for the GPT-6 family models Astra, Sol and Terra rise from 5,000 RPM and 1 million TPM on Build to 15,000 RPM and 40 million TPM on Grow. The Luna model goes up to 30,000 RPM and 180 million TPM. Long-context requests have a separate limit, some model families share one limit, and the Batch API counts queued input tokens per model.
Source: OpenAI, "Rate limits" guide (developers.openai.com/api/docs/guides/rate-limits), read 27 October 2026. Standard limits; Ultrafast mode has separate limits. RPM: Astra, Sol, Terra 5,000 / 10,000 / 15,000; Luna 5,000 / 10,000 / 30,000. OpenAI's TPM counts input and output together, Anthropic counts them separately, so the numbers are not directly comparable.
Google, Mistral and Manus
- Google Gemini publishes the tier rules and batch limits (Tier 1: 5 million enqueued batch tokens for Gemini 4 Argon, Tier 2: 500 million), but the per-model RPM and TPM numbers are shown only in Google AI Studio. Google states that the limits are not guaranteed and capacity may vary. Priority inference gets 0.3 times the standard limit.
- Mistral measures requests per second, tokens per minute and tokens per month, with input and output counted together. The free tier is meant for evaluation and prototyping; production workloads need a paid Scale plan. Mistral shows the actual numbers only in the Limits section of its console.
- Manus limits its agent API per user and per endpoint, the same for every plan: 10 new tasks or follow-up messages per minute, 40 updates or stops and 100 status reads per minute, 600 usage queries per minute. All API keys of one user share the counter. For an agent that runs for minutes per task, 10 starts per minute is rarely the bottleneck; credits are.
Prompt caching: the cheapest way to more Claude throughput
Anthropic's limits look lower than OpenAI's at first glance, but there is an important difference: for almost all current Claude models, tokens read from the prompt cache do not count toward ITPM. Only uncached input and tokens written to the cache count. With a 2 million ITPM limit and an 80% cache hit rate you can process about 10 million input tokens per minute, and the cached tokens are also billed at a fraction of the normal price.
Blocksize calculation from Anthropic's rule that cache reads do not count toward ITPM (most models): effective input tokens per minute = 2,000,000 / (1 − cache hit rate). Anthropic's documentation gives the 80% example (10 M). Output tokens (OTPM) still apply.
Applications with a long, stable system prompt, tool definitions or a large reference document gain the most. The same rule makes max_tokens harmless: Anthropic counts only output tokens actually generated, so a high max_tokens value does not consume OTPM in advance. See our guide to prompt caching for the setup.
Reading the 429: what to retry and what not
| Response | Cause | What to do |
|---|---|---|
429 with retry-after | RPM, ITPM, OTPM or TPM exceeded | Wait the stated seconds, then retry with exponential backoff and jitter |
| 429 after a traffic jump | Acceleration limit (Anthropic) or "slow_down" (OpenAI) | Ramp traffic up gradually; spread batch starts |
429, Anthropic enforced_spend_limit_reached | Monthly tier cap reached | Do not retry; request a tier increase |
| 400, "specified API usage limits" | Your own spend limit (Anthropic) | Raise or remove the limit in the Console |
| 503 "server_is_overloaded" (OpenAI) / 529 (Anthropic) | Temporary overload at the provider | Retry with backoff; consider a fallback model or region |
Both Anthropic and OpenAI return the current state in response headers (anthropic-ratelimit-input-tokens-remaining, x-ratelimit-remaining-tokens and their siblings). Log them. They show how close you run to the limit long before the first 429, and they are the best evidence when you ask for a higher tier.
Six rules for production
- Know your tier before launch. Check the limits page in the Console or AI Studio, and request an upgrade weeks before a campaign, not on the day.
- Separate workloads. Use Anthropic workspaces or OpenAI projects with their own limits, so a batch job or a test cannot starve the customer-facing chatbot.
- Cache everything that repeats. System prompts, tool definitions and long documents in the cache raise Claude throughput and cut costs.
- Move non-urgent work to batch. The Batch APIs have separate limits and half the price (see batch processing).
- Route by task. Small models have their own buckets; sending simple requests there frees capacity for the large model (see model routing).
- Back off properly. Honour
retry-after, use exponential backoff with jitter, limit concurrency in a queue and never retry a spend-cap error in a loop.
What this means for companies in Croatia and the DACH region
Most European SMEs start on the lowest tier, where the monthly cap ($500 at Anthropic and OpenAI Build, $250 at Gemini Tier 1) is reached faster than expected once an internal assistant is rolled out to the whole company. Plan the tier upgrade as part of the go-live, keep production and experiments in separate workspaces, and design caching from day one: it is the one measure that lowers cost and raises capacity at the same time. For critical processes, a second provider or a cloud route (Claude via AWS Bedrock or Google Cloud, with their own quotas) protects against a single 429 stopping your business.
Frequently asked questions
What are the Claude API rate limits in 2026?
Anthropic places organisations on Start, Build, Scale or Custom tiers. For Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 the standard limits are 1,000 requests, 2 million input and 400,000 output tokens per minute on Start; 5,000 requests, 5 million input and 1 million output tokens on Build; and 10,000 requests, 10 million input and 2 million output tokens on Scale. Claude Fable 5.x has lower limits. New organisations may start in an Evaluation tier below these numbers.
What are the OpenAI API usage tiers?
OpenAI now has three paid tiers. Build requires $5 in total credit purchases and allows $500 a month, Launch requires $100 and allows $5,000 a month, Grow requires $500 and allows $200,000 a month. Standard limits for the Astra, Sol and Terra models rise from 1 million tokens per minute on Build to 40 million on Grow; Luna goes from 2 million to 180 million.
Do cached tokens count toward Claude rate limits?
For almost all current Claude models, no. Only uncached input tokens and tokens written to the cache count toward the input-tokens-per-minute limit; tokens read from the cache do not. With an 80% cache hit rate, a 2 million ITPM limit therefore lets you process about 10 million input tokens per minute.
Why does the API return 429 even though I am under my rate limit?
There are three common reasons: a short burst, because a per-minute limit is enforced in smaller intervals (60 RPM can mean one request per second); an acceleration limit after a sudden jump in traffic; or a monthly spend cap. Anthropic's spend-cap 429 has no retry-after header and the error code enforced_spend_limit_reached, so retrying does not help until the next month or a tier upgrade.
Is your AI application ready for real traffic?
We review tiers, limits, caching and retry logic, separate workloads and plan the capacity for your go-live. As a Claude Certified Architect from Zagreb, we work with companies in Croatia and the DACH region.
Talk to an AI consultant