The conclusion first: as of October 2026, xAI has not published a free API tier that can be treated as a long-term commitment. The official API is billed by token, starting with the first request. What the market calls a "Grok free quota" is usually promotional credit, a third-party free model, a shared or resold key, or free chat in the web app or X app. Their limits and risks are completely different.

This article separates several kinds of "free", explains where the quota comes from, rate limits, data terms, and suitable use cases, then gives a seven-day trial evaluation method and a path to paid service. Specific prices are subject to model pricing.

Note: xAI and Grok are trademarks of their respective owners. This article is not affiliated with xAI. Free-tier policies, promotional credit, and product terms may change; the official console and terms are authoritative.

1. Four kinds of "free" are not the same

Type Programmatic access Quota source Main limits Risk
Grok web app or X app No In-product allowance Usage and feature caps Cannot be automated; consumer terms apply
xAI promotional credit (if any) Yes Official campaign One-time, time-limited, conditional Expires when the campaign ends
Third-party free model Yes Platform subsidy or limited-time event Restricted models and rates; possible queues Different terms and data path
Shared or resold key Nominally yes Someone else's account Rate-limited or banned at any time Leakage, compliance, and stability risk

For the first two, read the official rules. For the third, read the platform terms. The fourth is not recommended for production business. When you see "permanently free" or "unlimited calls", assume it can disappear at any moment.

2. Official free quota: what is visible as of October 2026

xAI's pricing and documentation pages list charges by token. They do not list a long-term free tier, nor do they publish a sign-up credit amount. Whether new-user credit or a limited-time promotion exists is subject to console announcements and terms; do not treat a third-party screenshot as policy.

Tier 0 is easily misread. It is the default tier for zero cumulative spending, not a free quota. Calls are still billed by token; the rate cap is simply lower. Tiers rise with cumulative spending, and limits vary by model and tier, with exact figures in official documentation and the console.

xAI has run promotional credit campaigns in the past, some related to data sharing or new model promotion. Promotional credit usually has an expiry date, eligibility rules, and usage restrictions, and it may end at any time. Putting promotional credit into a production budget is a common misjudgment.

Also distinguish consumer products from the API. Free chat in the web app or X app does not mean you can obtain an API key; accounts, billing, and terms are managed separately.

3. Five common limits of free tiers

  1. Rate and concurrency: Usually limited by requests per second and tokens per minute, with 429 returned above the limit. Trial or low-tier accounts often see more queuing and failures at peak times.
  2. Model range: Free models may be smaller, older, or available only for a limited time, with different context length and reasoning capability from flagship models.
  3. Feature cuts: Tool calling, structured output, file upload, batch processing, and long context may be unavailable or carry separate quotas.
  4. Data terms: Retention, training, and deletion rules can differ completely between the official API and a third-party free platform. When customer data, contracts, or code are involved, read the terms first.
  5. Service assurance: Free or trial services usually have no SLA, compensation, or reliable incident notices; rate limits, retirement, and rule changes may come without advance notice.

Any one of these can turn "free" into a more expensive engineering cost. The time spent on retries, migration, and compliance review is often higher than the API fee itself.

4. Which workloads are suitable for a free or trial tier first

Suitable scenarios usually share three traits: low call volume, acceptable failure, and non-sensitive data.

  • Learning courses, technical validation, and product prototypes.
  • Low-frequency internal scripts, such as generating a weekly report summary.
  • Assistive tasks whose output can be reviewed by a person.
  • Classification, rewriting, and summarization tests on public or de-identified data.
  • Short evaluations with a clear stopping condition, such as seven days or a fixed budget.

Unsuitable scenarios are equally clear: customer-facing production services, personal information or contract text, high-concurrency real-time calls, long-running batch jobs, and procurement that requires invoices and contracts. The test is simple: if a single failure means "try again later", the trial can continue; if a failure causes customer complaints or breach of contract, design for a paid plan.

5. Run a reproducible seven-day trial evaluation

Do not draw conclusions from a single conversation. Record the following to judge whether free or trial quota is sufficient:

  1. Draw 20 to 30 tasks from real business and define acceptance criteria; do not use prompts whose correctness cannot be judged.
  2. Run one round each morning and evening, recording status codes, latency, the actual model name returned, and output quality.
  3. Count 429s separately, distinguishing minute-level rate limits, daily quota exhaustion, and platform queuing.
  4. Record input and output tokens per task and convert to cost per task at current prices.
  5. Simulate a degradation plan for exhausted quota and confirm whether the business can accept it.
  6. Compare with a paid model or alternative model on quality, latency, and total cost.
  7. Write the conclusion as numbers: success rate, P95 latency, cost per task, and sustainable daily call volume.

Below is a trial script with a budget cap. It stops when the self-set token limit is reached, avoiding overspend during testing:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["GROK_API_KEY"],
    base_url=os.environ.get("GROK_BASE_URL", "https://ai.zhongshangqihang.com/v1"),
    timeout=60,
    max_retries=0,
)

BUDGET_TOKENS = int(os.environ.get("BUDGET_TOKENS", "200000"))
used = 0

def ask(prompt: str) -> str:
    global used
    if used >= BUDGET_TOKENS:
        raise RuntimeError("Trial budget exhausted; stopping calls")
    resp = client.chat.completions.create(
        model=os.environ["GROK_MODEL"],
        messages=[{"role": "user", "content": prompt}],
        max_tokens=512,
    )
    if resp.usage:
        used += resp.usage.total_tokens
        print(f"Tokens used: {used}/{BUDGET_TOKENS}")
    return resp.choices[0].message.content

print(ask("Explain in three sentences which metrics a free API trial should record."))

BUDGET_TOKENS is a team-defined cap, not an official quota. The script only addresses budget overrun; it does not solve rate limits or data compliance, which still require separate acceptance checks.

6. From trial to paid: three upgrade paths

  • xAI official pay-as-you-go: Suits teams with stable network access and an overseas payment method. Billing is in USD; account tier rises with cumulative spending, and rate caps are calculated per model. A higher tier is not a capacity commitment, and 429s can still occur at peak times.
  • Domestic AI gateway: Offers an OpenAI-compatible endpoint, RMB top-up, and unified billing, suiting teams that need invoices or multi-model switching. Before connecting, confirm model range, billing rules, data retention, and incident response; see model pricing for details.
  • Hybrid approach: Official direct access as primary with a gateway as backup, or a gateway as primary with official access as backup. Monitor both channels separately and rehearse switching regularly so no single entry point becomes a failure point.

Do not switch all traffic at once during upgrade. Top up a small amount, run real tasks for a week, then decide whether to expand based on success rate, latency, billing, and support response.

7. Security and compliance baseline

  1. Do not put keys in front-end code, Git repositories, or chat logs.
  2. Use one key per project, with budgets, quotas, and alerts set per project.
  3. Do not use "free keys" of unknown origin, especially channels that ask you to upload account documents or identity information.
  4. Before sending customer data, contracts, or code, confirm whether data is retained, whether it is used for training, and for how long.
  5. Keep usage logs and error records so you can determine whether a problem is quota, network, or model related.
  6. Prepare a service-stop plan: a backup model, a backup channel, and a degradation notice for users.

The official API provides one layer of data commitments; a third-party relay has its own logging and retention policy. Confirm both layers separately and do not treat upstream terms as the relay's terms.

Frequently asked questions

Does the Grok API have a permanently free tier?

As of October 2026, the official documentation does not list a long-term free tier. Whether promotional credit or a limited-time campaign exists is subject to console announcements and terms; do not treat third-party marketing as official policy.

Is Tier 0 a free tier?

No. Tier 0 is the default tier for zero cumulative spending. Calls are still billed by token; only the rate cap is lower. Tiers and free quota are two different concepts.

Can I use a Grok free key that someone gives away online?

It is not recommended for production business. A key of unknown origin may be shared, resold, or banned at any time, and request content may pass through a third party. Even for testing, use a separate deletable key and never send sensitive data.

Can a free API be used commercially?

Commercial use depends on official or platform terms and on the specific scenario. When customer data, external commitments, or invoices are involved, use a paid plan with clear commercial terms and billing. Do not assume that "it can be called" means "it can be used commercially".

Is the free quota the same model as the paid quota?

Not necessarily. Free models may be smaller or older versions, and context and features may be cut. Check the actual model identifier returned, not just the promotional name.

Next step

Free or trial quota is suitable for validating an idea, not for promising production service. Spend seven days obtaining success rate, latency, and cost per task, then decide whether to stay on the trial, move to official paid service, or use a unified gateway. For RMB billing and a multi-model entry point, create a test key at ZSQH AI Gateway; prices are subject to model pricing.

Related reading: