Free tiers & deals
These are the no-charge notebooks, shared graphics chips, and API caps you can use today, and the rule for when a logged price drop counts as a deal. The offer market opens later.
Beginner
Use a free allowance before you pay
A free tier is a no-charge allowance: a browser notebook, a short turn on a shared graphics chip, or a daily cap on API calls. An API sends text to a model and returns text, counted in tokens, which are small pieces of words.
Google Colab hosts a Jupyter notebook (notes plus code) and can attach a GPU, a graphics chip used for neural networks, or a TPU, Google's accelerator, free of charge. Access is not guaranteed, chip types change, and Colab does not publish the limits. A free notebook runs for at most 12 hours. Idle runtimes stop. Free runtimes disallow remote desktops, SSH, and using the machine as a website.
Static Spaces on Hugging Face are free. ZeroGPU lends an NVIDIA RTX Pro 6000 only during a marked function: 2 minutes of GPU time a day signed out, or 5 minutes on a free account, resetting 24 hours after the first use. The default slice is half the card, 48 GB of video memory.
OpenRouter's free plan lists 25+ models with no token price: 20 requests a minute and 50 a day, or 1,000 a day after the account has bought at least $10 of credits. A negative balance blocks free models. Read the provider's data policy before you paste private text.
If you are 13 or older and enrolled in a degree- or diploma-granting program (high school, college, university, or homeschool), apply for GitHub Education with a personal account and a current school ID, class schedule, transcript, or enrollment letter.
Plan my first free AI setup
I am an adult learning AI. Ask me three questions: am I currently enrolled in a degree or diploma program, can I use a Google account, and may I install software on this computer? Then give me an order for this week using only these pages, and quote each limit from the page: Google Colab FAQ (https://research.google.com/colaboratory/faq.html), Hugging Face ZeroGPU (https://huggingface.co/docs/hub/spaces-zerogpu), OpenRouter pricing (https://openrouter.ai/pricing), and GitHub Education (https://docs.github.com/en/education/explore-the-benefits-of-teaching-and-learning-with-github-education/github-education-for-students/apply-to-github-education-as-a-student). If a limit is unpublished, say so. Do not invent hour quotas. End with a five-line log: date, tool, GPU name shown, what I ran, when I disconnected.
Intermediate
Score a price against a log
Polaris will log posted text, image, and video API prices once a day. For the first 30 days the log only collects rows. A price is marked as a deal only after that seller and product have 30 days of history, and only when the new price is under both the 30-day low and the lowest price ever logged, with the ordinary price and the seller shown beside it. A sale badge with no history fails that test.
OpenRouter's pricing page links a public models feed, no key required. Each model has a pricing object for prompt and completion tokens in US dollars. LiteLLM publishes a JSON catalog of prices and context windows. A context window is how much text one request can hold. Copy the date, model id, seller, input price, output price, and context length. When a number changes, add a row and keep the old row.
Colab does not name the free GPU or a weekly hour quota, so record the accelerator the session actually reports. Hugging Face publishes Spaces prices: CPU Basic has no hourly fee, an Nvidia T4 small is listed at $0.40 an hour with 16 GB of GPU memory, and an Nvidia A100 large at $2.50 an hour with 80 GB.
Use a free GPU for one saved experiment. Load one open model, run a few prompts, export the output, and disconnect. A spreadsheet with those columns works today.
Test this price against my rows
I will paste rows with date, seller, product, input price, and output price. Apply only this rule: a row is a deal if that seller and product already have 30 days of rows, and the newest price is below both the lowest price in those 30 days and the lowest price in the whole log. If there are fewer than 30 days, say collection only and do not call it a deal. Show the ordinary price and the seller. Then name the fields to copy from https://openrouter.ai/api/v1/models and from https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json. Here are my rows:
Expert
When a smaller copy beats a rented API
Quantization stores each learned weight in fewer bits. At 16-bit, one weight is two bytes, so 8 billion parameters are about 16 GB of weights before prompt memory. A 4-bit file is about a quarter of that and may fit a smaller graphics card. Keep the smaller file only if the same questions still score well enough for your task.
Multiply one real week of input and output tokens by the prices in your log. Compare that cost with electricity and time for the local file. If the weights fit and the API week costs more, local is the cheaper path for that job.
Google for Startups Cloud gives credits over two years. Other startups get up to $100,000 in year one, and AI startups up to $250,000. In year two Google covers 20 percent of spend, up to another $100,000, so an AI startup can reach $350,000. You need a Google Cloud account with a billing ID, a company email that matches your public website, Pre-Seed to Series A funding, and a founding date within the last three to five years, depending on the tier.
NVIDIA Inception has no fee and takes no equity. The company must be incorporated, under ten years old, have a working website, a founder running it, and a clear NVIDIA use. Consulting and outsourced development firms, cryptocurrency companies, cloud providers, resellers, distributors, and public companies are excluded. Hardware discounts are a supplier rebate. Cloud credits from NVIDIA and partners have no public dollar amount. Apply once per company. Colab bans extra accounts made to dodge a quota.
Compare a local file with my API bill
Ask me for parameter count, bit width, GPU memory in GB, input tokens per week, output tokens per week, and the per-token prices from my own log. Compute weight memory as parameters times bytes per weight (16-bit = 2, 8-bit = 1, 4-bit = 0.5). Say whether the weights alone fit, and say that the prompt and the running calculation need memory beyond that figure. Compute the weekly API cost. Then list the eligibility lines I must re-read on https://startup.google.com/cloud/ and https://www.nvidia.com/en-us/startups/ before I apply. Do not state a credit amount that I did not paste from those pages.
Check the offer before you start
- Write the date, the seller, and the ordinary price beside any discount.
- See whether a card is required, and what the price becomes when the free period ends.
- Read the host's data policy before you paste private text, code, or customer records.
- On Colab, set Hardware accelerator to None when the code is not using a GPU, and close the notebook when you stop.
- Use one account. Colab disallows extra accounts made to get around a limit.
- Students apply for GitHub Education with a personal account and current proof of enrollment. Colab's free one-year Pro for Education plan is closed to new signups.
- For startup credit, use a company email on the same domain as the public website, and send one application.
Which path fits the job
| Path | Use it for | Limit you can quote |
|---|---|---|
| Colab, free | Learning in a notebook | At most 12 hours; chip and other limits unpublished |
| Hugging Face ZeroGPU | A short demo | 2 minutes a day signed out, 5 minutes on a free account; default slice 48 GB. A verified account older than 30 days can host up to two ZeroGPU Spaces of its own |
| Hugging Face Static Space | A page with no GPU | Free for everyone |
| OpenRouter free plan | A few API tests | 20 requests a minute and 50 a day, or 1,000 a day after $10 of credits purchased |
| GitHub Education | Student tools | Age 13 or older, current enrollment, personal account |
| Google for Startups Cloud | A startup cloud bill | Up to $200,000 over two years, or up to $350,000 for AI startups, paid as cloud credits |
| NVIDIA Inception | Startup membership | No fee and no equity; the public page states no credit amount |
Re-open the linked page on the day you rely on a number. These allowances move.
How a deal gets listed
The daily log stores date, seller, product, ordinary price, observed price, and context length for text, image, and video APIs. For the first 30 days it only collects rows. After that, a price is listed as a deal only when it is under both the 30-day low and the lowest price ever logged for that seller and product, and the ordinary price and the seller are printed beside it. A discount banner with no history stays off the list. No deal is shown until that log exists.
A shared community DGX Station would take jobs that a free notebook stops mid-run, so a week of API spend can be set beside what the same job costs on the shared machine if it is bought; Hrafninn Industries AB would own it. About the shared station
Sources
- Google Colab FAQWhat the free notebook includes, the 12-hour cap, unpublished limits, and uses Colab disallows, including extra accounts and remote desktops.
- Hugging Face Spaces overviewStatic Spaces are free, free hardware sleeps, and the hardware table lists CPU Basic at no hourly fee, T4 small at $0.40 an hour, and A100 large at $2.50 an hour.
- Hugging Face Spaces ZeroGPUDaily GPU minutes by account type, the 48 GB and 96 GB slices of an RTX Pro 6000, and who may host a ZeroGPU Space.
- OpenRouter pricingThe free plan's model count, the 20-per-minute and 50-per-day caps, the $10 credit threshold, and the link to the public models price feed.
- LiteLLM model prices and context windowsA public JSON catalog with per-token prices and context lengths you can copy into a price log. The file is LiteLLM's data; use it as a feed, not as a reprint.
- Apply to GitHub Education as a studentWho qualifies, which enrollment documents are accepted, and where the application starts.
- Google for Startups Cloud ProgramThe published credit ceilings for early, funded, and AI-first startups, plus the website, email, and age requirements.
- NVIDIA Inception for StartupsFree membership rules, who is excluded, hardware rebates, and cloud credits offered with no dollar amount on the page.