AIVAX

Plans and Limits

AIVAX has four account plans: Free, Pro, Max, and Ultimate. The current plan controls model access, account and product rate limits, storage quota, retention periods, and included daily service allowances. Every plan limit is listed on this page; product pages link here instead of repeating the values.

For commercial subscription prices and plan packaging, use the AIVAX pricing page. The same public list prices are available from Plan list prices.

An em dash (—) means the plan does not impose a limit, and Not available means the feature is unavailable on the plan. Model, gateway, provider, or product-specific limits may still apply. Limits that are the same on every plan, such as file sizes or results per request, are documented on each product page.

Account limits #

LimitFreeProMaxUltimate
Inference
Model accessLow-price modelsAll modelsAll modelsAll models
Inference commission multiplier1.25x1.05x1.02x1.00x
Daily inference commission ceiling$1.50$1.20$1.00No commission
Integrated model requests20/min and 500/day150/min1,000/min—
Integrated model input tokens1,000,000/min15,000,000/min60,000,000/min—
BYOK requests30/min150/min1,000/min—
Maximum context64,000 input tokens———
Monthly subscription model quotaNot availableIncludedIncludedIncluded
Account and support
Storage quota30 MB2 GB10 GB20 GB
Cost per excess GBNot available$0.50/GB/month$0.30/GB/month$0.20/GB/month
Conversation retention2 hours2 days14 days30 days
RAG search history retention2 hours3 days14 days30 days
Service fee on credit top-ups5% (minimum $0.20)NoneNoneNone
Support levelEmailPriorityPriorityDedicated

Document text stored in a collection counts toward the storage quota as UTF-8 bytes. Text with accents or non-Latin characters can use more of that quota than basic Latin text of the same length. RAG search history is the list of searches shown in RAG Observability.

Integrated model requests are limited by both request count and input tokens. Model rate-limit groups adjust the request-count thresholds:

Rate-limit groupThreshold multiplier
Common1.0x
Discounted0.5x
Low0.3x
Free0.1x

For example, a Pro account normally has 150 integrated-model requests per minute. With a Discounted model group, the adjusted threshold is 75 requests per minute.

BYOK uses a provider key configured on the gateway instead of an integrated AIVAX model, but requests still pass through AIVAX infrastructure and use the plan’s BYOK limit.

Rate limits are shared across all API keys of the same account. A request that would exceed an active window returns 429 Too Many Requests; pace requests and use bounded retries with backoff.

Custom accounts use the Pro plan rate limits multiplied by a factor agreed for the account, or have no plan rate limits when no factor is set. Custom accounts support 8 concurrent Agentic Test runs.

Product limits #

Product limits use the same rules as account limits: they are shared across the account’s API keys, and a request above an active window returns 429 Too Many Requests. Where a limit counts items instead of HTTP requests, the note below its table says so.

RAG and collections #

LimitFreeProMaxUltimate
Collections51001,000—
Documents per account2,000200,000——
Documents per collection1,00050,000250,000500,000
Searches20/min300/min1,500/min3,000/min
Reranking requests30/min500/min3,000/min—
Reflex processing time30 minutes/day3 hours/day12 hours/day—
Document insertions500/day5,000/day50,000/day—
Documents per multipart JSONL import1,0005,000100,0001,000,000
Documents per WebSocket JSONL import10,00050,0001,000,00010,000,000
Media Injector files2/day20/day250/day1,000/day

The reranking request limit applies to the reranking endpoint and to RAG searches that use a reranker, including searches through AI Gateways and MCP tools. The Reflex processing-time limit counts the time spent processing Reflex requests on both paths; cached input does not consume it separately. If an import exceeds the per-import limit, split it into multiple imports. Document limits count stored documents and are checked only when a new document is created; updating or reindexing an existing document is not blocked. Once a limit is reached, creating a document returns 400 Bad Request and an import stops with an error. The count can take up to 30 seconds to reflect recent insertions or deletions. See Collections, Documents, and Reflex.

Data processing #

LimitFreeProMaxUltimate
Text-classification documents30/min and 300/day500/min2,500/min10,000/min
Extraction, summarization, and instruction refinement documents10/min and 300/day100/min1,000/min2,500/min
Text-segmentation documents10/min and 100/day300/min1,000/min2,500/min
Translation requests30/min and 300/day100/min1,000/min2,500/min
Media description requests10/min and 30/hour40/min100/min200/min
Semantic decision requests10/min30/min200/min—

Classification, extraction, summarization, and segmentation limits count every item in the request’s documents array, not each HTTP request; extraction and summarization share one quota. Instruction refinement shares that quota and counts every item in its context array. Translation, media description, and semantic decision limits count HTTP requests. A semantic decision request consumes one unit regardless of how many questions it contains or which decision model it selects, and included decision usage still consumes request units. See Classification, Extraction, Summarization, Instruction Refinement, Text segmentation, Translations, Media descriptions, and Semantic Decisions.

Speech #

LimitFreeProMaxUltimate
Standalone text-to-speech requests3/min and 40/hour20/min100/min300/min
Standalone audio-transcription requests3/min and 40/hour20/min100/min300/min

Text-to-speech and transcription have separate quotas with the same values. Voice Sessions are not subject to these standalone quotas. See Text-to-speech and Speech-to-text.

LimitFreeProMaxUltimate
Web search15/day500/day4,000/day10,000/day
X/Twitter searchNot available200/day3,000/day10,000/day
Document and web page generation5/day500/day15,000/day50,000/day
Image generation and editing5/day150/day1,500/day5,000/day
General service actions30/day2,500/day30,000/day100,000/day
Bash commands300/hour30,000/hour100,000/hour—

General service actions share one daily quota. The image generation limit is one daily quota shared by the built-in tool, the image generation API, and the Media generation MCP; each request counts once, whatever the number of images it asks for. See Built-in tools, Web Search, and Shell.

Batch and Agentic Tests #

LimitFreeProMaxUltimate
Batch workflow items processed500/day25,000/day250,000/day—
New Agentic Test runs5/min20/min100/min—
Concurrent Agentic Test runs1468

If batch processing is paused because of quota, it resumes after the window resets. Manual runs, scheduled runs, and direct evaluations share one new-run quota. See Batch and Agentic Tests.

Included daily subscription allowances #

Every plan includes separate daily allowances for the services below. Each comparison refers to the same service on the Free plan, not to a shared credit balance or a guaranteed number of requests. Unused allowance from one service cannot cover another. Custom accounts do not receive subscription allowances.

Included serviceFreeProMaxUltimate
RAG search and insertion embeddings10,000 tokens15× Free50× Free100× Free
Reranking with Reflex2,000 input tokens4× Free20× Free50× Free
Semantic decisions with eligible models20,000 input tokens2× Free3× Free5× Free
Fetch and OCR extraction1,000 PUs8× Free25× Free50× Free

RAG searches and document insertions share the embedding allowance. It does not cover answer generation, media processing, document extraction or summarization, text classification, or segmentation. A query embedding served from cache does not consume it. Reflex uses a separate reranking allowance that includes both cached and uncached input. The semantic decision allowance covers only eligible decision models; see Semantic Decisions. Optional Fetch JSON conversion is separate from the extraction allowance.

Coverage is evaluated for each metered service item: a document’s embedding, an individual query-term embedding, a reranking call, a decision call’s input usage, or an extraction operation. Each item is either fully included or billed in full at normal rates. Included items are tracked in subscription consumption, not as zero-cost entries in billing history. The current allowances permit a 10% margin above their base capacity. An item that would exceed that margin leaves the allowance unchanged and is billed normally. One request can contain several items, so some may be included while others are charged.

Daily allowances reset at midnight in the server’s local time. Check the account’s subscription usage indicators for consumption and reset status; usage can exceed 100% while within the margin. These daily allowances do not cover text-model inference; subscription models use a separate monthly quota, and inference with other models is metered against the account balance. Allowances do not bypass balance requirements or rate limits. See Pricing for how uncovered items are billed. For a walkthrough of testing a retrieval pipeline within these allowances, see Free RAG API on AIVAX Free.

Monthly subscription model quota #

Pro, Max, and Ultimate include a monthly quota for subscription models. Free and Custom plans do not include it. The quota window starts with the first request that uses it and lasts one month, so it renews on the monthly anniversary of the first use, not on the 1st. A month without any use starts no window. All subscription models share it. Max includes 5× the Pro quota, and Ultimate includes 12.5× the Pro quota. For the available models and the usage each plan includes, see Subscription models.

Subscription models are the catalog models whose names start with @aivax-subscription/. Using them never debits the account balance, and they have no per-token prices. Every other model is metered against the balance as usual. To tell them apart, check flags.isSubscription in GET /api/v1/information/models.json or _details.provider_model.subscription in GET /v1/models.

When the quota is exhausted, or the plan has none, a request to a subscription model returns 429 Too Many Requests with a message that the subscription model quota is exhausted. The request is not routed to a model outside the subscription. On plans with the quota, requests succeed again after the window renews one month after the first use; until then, select another model or move to a plan that includes it.

The rate at which each subscription model consumes the quota is dynamic and can change daily. See Consumption rate.

The account balance response reports the quota only as a used percentage, in a subscription consumption of type InferenceMonthly. planLimits.includesSubscriptionModels is true on plans that include the quota. See Get Account Balance.

Estimating PUs and tokens #

Usage is metered in processing units (PUs) or tokens, depending on the service. The examples below help you estimate the size of a workload. They are approximations (≈), not guaranteed values: the exact amount is reported in each response’s usage fields.

Fetch and OCR processing units

ContentApproximate PUs
Plain text or Markdown file0
Native-text PDF or DOCX, 10 pages≈ 1 (minimum of 1 PU per item)
Web page that renders in 1 to 3 seconds≈ 10–30 (10 PUs per second of rendering)
Photo or screenshot with text≈ 40–50
Scanned PDF, 10 pages≈ 400–500

Native pages are converted from their text layer and cost almost nothing; scanned pages, images, and images embedded in a page go through OCR and drive the total. A PDF that mixes both is billed for the pages that actually need OCR.

Tokens

ContentApproximate tokens
RAG search query, such as “how do I reset my password if I lost access to my email?”≈ 10–30
One page of English text, about 500 words≈ 650–750
RAG document chunk of 1,500 characters≈ 350–400
Reflex reranking of 50 candidates with about 300 tokens each≈ 15,000–16,000, since the query is scored with every candidate
A 300-page book, about 300 words per page≈ 110,000–130,000

Token estimates assume English text, where one token is about four characters or three quarters of a word. Portuguese, Spanish, and other accented languages usually need 20–50% more tokens for the same content; code, JSON, and non-Latin scripts can need more. A query embedding served from cache is not billed, and in chat completions the conversation history and system instructions are sent again on every turn.

Public API keys #

Public keys have additional limits independent of the account plan.

ScopeRequest limits
Per remote address3/5s, 20/min, 300/hour, 1,000/day
Global per key10/5s, 60/min, 1,500/hour, 10,000/day
ScopeToken limits
Per remote address100,000/5min, 500,000/30min, 2,000,000/6h, 5,000,000/day
Global per key500,000/5min, 2,000,000/30min, 10,000,000/6h, 25,000,000/day

Public keys can be used for chat completions, the /api/v1/generations/* endpoints, RAG semantic search, and RAG answer generation. For chat completions, public keys also require a full AI Gateway UUID, restrict request parameters, and omit server-side tool surfaces. See Authentication.

Type to search the documentation.