# xAI rate limits

Status: DOCUMENTED (tier tables) + LIVE_DISCOVERED (undocumented x-ratelimit-* response headers observed with our Tier 0 key on 2026-09-19). No 429 was triggered.
Sources: https://docs.x.ai/developers/rate-limits · /developers/models/ (per-model "Rate limits" tables = Tier 0 values) · /developers/management-api-guide (per-key qps/qpm/tpm) · /developers/debugging · /developers/advanced-api-usage/batch-api · tmp-live/xai-models/catalogue-probe.json.
Last verified: 2026-09-18 · Machine-readable: generated/fragments/rate-limits/xai-rate-limits.json.

# 1. Principles

Limits are per team, per model, on two dimensions: RPS (requests per second, derived from a per-minute request budget: RPM/60, so a minute's budget cannot be spent in one second) and TPM (tokens per minute).
Tier = cumulative spend on the xAI API since 2026-01-01 (prepaid purchases or fulfilled invoices). Automatic, permanent (never downgrades).
TPM counts prompt tokens (text, image, audio), completion tokens, reasoning tokens and cached prompt tokens.
Exceeding either → HTTP 429 (gRPC RESOURCE_EXHAUSTED). Docs recommend exponential backoff (2**attempt, 5 retries). No Retry-After documented.
Batch API requests do not count towards rate limits. Tiers apply to text, embedding and voice models; Imagine (image/video) limits are RPS-only and increased via sales@x.ai.
Per-key caps (in addition to team limits): qps, qpm, tpm set when creating/updating a key through the Management API (POST /auth/teams/{teamId}/api-keys, PUT /auth/api-keys/{id} with fieldMask). The tpm limiter engages when strictly exceeded and does not abort in-flight requests.
Personalised limits are on the console Models page (https://console.x.ai/team/default/models, ?cluster=us-central-1 for the US endpoint).

# 2. Tiers

Tier	Cumulative spend
Tier 0	$0 (default)
Tier 1	$50
Tier 2	$250
Tier 3	$1,000
Tier 4	$5,000
Enterprise	on request

# 3. Per-model limits (RPS / TPM by tier)

Model	RPS T0 · T1 · T2 · T3 · T4	TPM T0 · T1 · T2 · T3 · T4
grok-4.6	150 · 172 · 208 · 312 · 500	50M · 53M · 60M · 74M · 100M
grok-4.5	150 · 172 · 208 · 312 · 500	50M · 53M · 60M · 74M · 100M
grok-4.3	37 · 50 · 75 · 125 · 208	10M · 15M · 25M · 45M · 85M
grok-4.20-0309-reasoning	37 · 50 · 75 · 125 · 208	10M · 15M · 25M · 45M · 85M
grok-4.20-0309-non-reasoning	37 · 50 · 75 · 125 · 208	10M · 15M · 25M · 45M · 85M
grok-build-0.1	37 · 50 · 75 · 125 · 208	10M · 15M · 25M · 45M · 85M
grok-4.20-multi-agent-0309	9 · 12 · 18 · 31 · 56	2.5M · 3.7M · 6.2M · 11M · 21M
grok-imagine-image, -quality, -2.0	6 · 12 · 25 · 50 · 100	—
grok-imagine-video, -1.5	10 · 20 · 39 · 79 · 158	—

# Voice & audio

Service	RPS T0 · T4	Concurrent sessions T0 · T1 · T2 · T3 · T4	Other
grok-voice-think-fast-2.0 (speech-to-speech)	—	10 · 20 · 50 · 100 · 200	max session 120 min
Text to Speech	50 · 50 · 100 · 250 · 500	100 · 200 · 200 · 300 · 500
Speech to Text	10 · 10 · 20 · 30 · 40	100 · 200 · 200 · 300 · 500 (streaming)

Embedding models: tiers apply, no table published.

# 4. Observed headers (our key, 2026-09-19 — account-specific, not documentation)

Request	`x-ratelimit-limit-requests`	`x-ratelimit-limit-tokens`	remaining-*
POST /v1/chat/completions grok-4.6, grok-4.5	7200	50000000	equal to limits after one call
POST /v1/chat/completions grok-4.3, grok-4.20-0309-non-reasoning, grok-build-0.1	1800	10000000	idem
POST /v1/tokenize-text (grok-4.3)	1800	(absent)
POST /v1/responses grok-4.20-multi-agent-0309	(absent)	(absent)
GET /v1/models, /v1/language-models, /v1/api-key, /v1/me	(absent)	(absent)

Interpretation: x-ratelimit-limit-tokens equals the documented Tier 0 TPM. x-ratelimit-limit-requests looks like a per-minute request budget (7200 → 120/s, 1800 → 30/s), i.e. below the documented Tier 0 RPS (150, 37) — possibly a new-team allowance; the docs' "RPS = RPM/60" statement fits this header being the RPM. No x-ratelimit-reset-* header exists. Every inference response also carries x-request-id, Cloudflare Server-Timing and CF-RAY.

# 5. Increasing limits

Spend more (automatic), request an increase from the console Models page (also for limits beyond Tier 4), or contact sales@x.ai. For Imagine limits contact sales.

max_session_minutes 120 for speech-to-speech.
Deferred completion results are retrievable once within 24 h; batch completes typically within 24 h.
Prepaid credits exhausted with a $0 invoiced-billing limit → requests rejected (status code not documented).