2026 Free AI API Credits: Gemini, OpenAI, Claude, DeepSeek

Checked on: 2026-09-15 (Taipei time). Free plans change at any time. This article labels every cell with its source and check date; before applying, go by the official page you see at that moment. Where I could not verify a figure, this article does not state one.

There are many articles about "free AI APIs," but most mix together "has a free tier," "gives credits," "student plans" and "cloud credits" without stating the conditions. This one does one thing: it puts what each provider's free quota actually is, who it is for, what its limits are, and when it expires into one table, including the condition under which it does not apply to you.

The whole picture in one table (2026-09-15)

Three things that stand out at a glance:

1. Only the Gemini Flash free tier and the BazaarLink free tier are available to anyone who signs up with no card. Everything else requires a card (AWS), a payment first (OpenAI upgrades), being an existing paying customer (Anthropic), or simply does not exist (DeepSeek).
2. The Gemini free tier has a cost few people mention: its content is used to improve products. Content with confidentiality requirements cannot go through the free tier, regardless of how much quota there is.
3. Figures like "send $300" or "send $200" depend on who receives them: AWS is $200 not $300, and the Anthropic amount is a reward for existing subscribers.

Gemini API free quota: still there, but changed

This is the provider most people ask about, so it deserves its own section. The free tier still exists, but three changes shown on the official page as of 2026-09-15 are worth knowing:

• Flash series only. The 3.8, 3.7, 3.6 and 3.5 Flash, 3.5 and 3.1 Flash-Lite, and 2.5 Flash / Flash-Lite free tiers are all Free of charge; 3.1 Pro and 2.5 Pro are Not available. Gemma 4 is free.
• The RPM, TPM and RPD figures have been removed from the official docs. The rate-limit page now only says "depends on your usage tier, check in AI Studio." Articles online that say "N per day, M per minute" are either outdated or reflect one project's values at a given moment. To know your own quota, you can only log in to AI Studio and look. The community's talk of being "cut" or "shrunk" is mostly about this.
• Upgrading requires a Google Cloud billing account and a foreign-currency credit card. Free to Tier 1 takes effect immediately, with a monthly cap of $250; Tier 2 requires $100 cumulative payment and at least 3 days; Tier 3 requires $1,000 and at least 30 days.

Paid unit prices for each model, a New Taiwan dollar estimate, and three paths for when the free quota runs out are covered in Gemini API pricing and free quota 2026.

OpenAI and DeepSeek: what the official pages say

OpenAI: the rate-limit page lists six usage tiers. The "Free" tier's condition is that "the user must be in an allowed region," with monthly usage capped at $100; Tier 1 and above require a $5 payment first. The page nowhere mentions that new accounts receive free credits. So claims online that "signing up gives you credits" are not included in this article, because they cannot be found on the official page.

DeepSeek: the pricing page has no mention of free credits, trials or giveaways at all. But it has something practical: off-peak half pricing. Peak hours are UTC Mon–Fri 01:00–04:00 and 06:00–10:00 (Taipei time 09–12 and 14–18), and at all other times every model's price is halved. For batch jobs, this is far more useful than a "free quota."

AWS credits: the official figure is up to $200, not $300

Many articles say "AWS gives new accounts $300." On 2026-09-12, the AWS official free plan page said up to $200, in two parts:

• $100 USD on sign-up
• Up to another $100 USD for exploring the main AWS services
• The free plan lasts up to 6 months

Amazon Bedrock can be used on both the free plan and paid plans (Bedrock AgentCore is exclusive to paid plans). So "running Claude with AWS credits" is technically viable, but it has two practical pitfalls, covered below.

⚠️ Pitfall one: Claude on Bedrock is not cheaper

We pulled Bedrock's price list ourselves and checked it (2026-09-08, using aws pricing get-products): Claude on Bedrock is priced at Anthropic's official standard rate, exactly the same as buying directly from Anthropic, so there is zero savings. Among Bedrock's four billing tiers (standard / flex / priority / batch), the flex tier is not offered for Claude, so there is no discount to be had that way. The credits are real, but the unit price is no lower — you are just spending the $200 sooner.

⚠️ Pitfall two: Claude on Bedrock cannot be called with a simple Bearer key

In the same test, we also called Bedrock's OpenAI-compatible endpoint: the Anthropic models did not work with a Bearer key. /chat/completions and /responses both returned "model does not support this API," and /messages returned 404. Claude must go through Bedrock's native Converse API, which requires SigV4 signing; swapping the base URL is not enough. If your existing code uses the OpenAI-compatible format, there is more to change here than you might expect.

(Incidentally, in the same test 38 open-source models worked normally with a Bearer key, so this route is much friendlier to open-source models than to Claude.)

Anthropic's promotional credit: not for the general public

The "$200 Claude credit" that is often reposted requires that you are already a paying subscriber (Pro / Max / Team), and you must claim it within the announced deadline; late claims are not honored.

In other words, it is not an entry-level credit for newcomers; it is a reward for existing paying users. If you have never paid and want to try the Claude API for free, this route does not apply to you from the start.

Promotions like this also share two traits:

• Tied to subscription status: usually lapses along with the subscription when you cancel
• Time-limited: not a permanent balance; it goes to zero when it expires

Other common free credits and their real thresholds

The most commonly misunderstood row is student plans. Those plans give web or app subscription access, and do not give you an API key, nor can they be used to run your own programs.

So is there anything that needs no card and no upfront payment?

Yes, with two conditions: the model is fixed, and there is a daily limit.

Our free tier currently works like this (the numbers come from the platform's live settings, not marketing copy):

Two things to be clear about:

1. There is no sign-up credit. We do not give a dollar balance; we provide free usage of that model. (We once gave a very small sign-up credit, which stopped in August 2026.)
2. The free quota covers only that one model. If you want to use paid models such as Claude, GPT or Gemini, you pay list price; there is no free version.

To call it, set the model to auto:free, or specify that model directly:

curl https://api.bazaarlink.ai/v1/chat/completions \
  -H "Authorization: Bearer $BAZAARLINK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "auto:free", "messages": [{"role": "user", "content": "hello"}]}'

When the free quota runs out

This is the part most "free quota" articles leave out, but it is the problem you will actually face.

Credits run out, promotions expire, and free tiers have daily caps. These plans are meant to let you try things out, not to run a business long-term. So when evaluating, it is worth calculating one thing first: once you are live, roughly how much will you spend per month? If that number is small, the free tier may be enough; if it will exceed it, look at how regular billing works from the start, rather than spending time moving between providers' free quotas.

For the actual unit prices and New Taiwan dollar estimates for each provider's models, see the four price articles: Claude, GPT-5, Gemini and DeepSeek.

Once usage reaches a certain scale, part of it comes back

This is a separate matter from free credits, but it answers the same question: how much will it cost in the long run.

BazaarLink currently offers usage rebates for OpenAI GPT, Gemini and DeepSeek. Each time qualifying monthly spend crosses a threshold, a fixed amount is issued and goes directly into the account balance. Reached thresholds accumulate; you do not get only the highest tier.

But to be honest: this rebate is almost useless to someone who has just graduated from free credits.

Monthly spend under USD 20 earns 0, and projects that have just started paying usually spend single-digit dollars a month. The amount also varies: landing exactly on a threshold gives 10%, while spend between thresholds dilutes it (spending $199 returns only $10, which is 5%), and it only stabilizes at 10% above $500.

So the right order is: first build with the free tier, then estimate your monthly bill after launch using unit prices; rebates only start to matter once your monthly bill has risen. For the rules and the list of models that qualify, see the usage rebate explanation.

Three more things worth knowing: usage on the free plan does not count, BYOK (bring your own key) usage does not count, and rebates are settled by UTC calendar month and recalculated every month; they do not accumulate.

TWD billing · Taiwan invoices · leading AI models · OpenAI-compatible API