Today's find
Today's find is a short dated note on something you can use for less: a free tool, an open model, a lower API rate, or an hour of graphics-card rental. The note names the page it came from, when that page was opened, and when to check it again.
Latest finds
Beginner
Free tools you can open today
A useful first find is a browser tool you can open today, with the limit written beside the word free.
Google Colab runs notebooks in the browser. You type Python, press run, and the work happens on a machine Google provides for that session. The Colab FAQ says the service is free of charge and can include GPUs and TPUs. A GPU is a graphics chip used for model math. A TPU is Google's chip for that same work. Free GPU access is heavily restricted, limits stay unpublished because they change, and a free notebook can run for at most 12 hours, depending on availability and use. Idle notebooks shut down. You need a Google account. Quote a weekly hour-count only if that FAQ prints it.
Practical Deep Learning for Coders is free for people who already write a little code. Its Kaggle page says the account must be phone-verified before a Kaggle GPU will run. Read the signup page before you add a phone number.
On each card, write the full page address, the date and time you opened it, whether a card or paid plan is required to start, and any end date the page prints. If it prints none, say so and look again within seven days. A chat window, an API key and a rented GPU are three different kinds of offer, so check each one separately. An API key is a secret an app sends so a company will run a model for you.
The tip form opens later. Check the official page again before you pass a find on.
Check a free-tier page
Check one AI offer for someone new to this. Use only the page I paste. Make a table with these rows: what I can do there; whether the page says it is free of charge; whether a credit card is required to start (quote the page, or write not stated); any time limit, idle cutoff, or GPU restriction, quoted; what a paid plan adds; any date printed on the page. Then write a find card with six lines: name, full URL, checked at (use today's date and ask me for the clock time if you lack it), card or phone required, the limit, and recheck by. Leave a row as not stated when the page is silent. Do not fill gaps from memory. Page: [paste the page text or address]
Intermediate
Open models and the bill per million tokens
An intermediate find is a model, or a block of API credit, that might cover work you now pay a monthly fee for. The card shows the license and the bill.
A license says what you may do with the weights, the saved model, and the code. Apache 2.0 and MIT allow use, changes, and shipping in a product, if you keep the notices they name. Other Hugging Face tags are narrower. Llama, Gemma, and OpenRAIL-style names often add use rules or a form. Filter the model list to apache-2.0 or mit, sort by Trending, and open the model card and the LICENSE file. If they disagree, follow the file and write that on the card.
A score is separate from a price. The Arena leaderboard dataset stores a license next to a preference score. In the agent subset, latest split, published 19 October 2026, rows included MiMo-V2.6-Pro and DeepSeek-V4.1-Flash-20260910 (Max), tagged MIT, plus Hy4 preview and Qwen3.8-27B, tagged Apache 2.0. Those rows are votes on that board. Check price, privacy, and your own task as well.
Vendors quote price per million tokens. A token is a chunk of text, often part of a word. Input is what you send. Output is what comes back, and the rates often differ. Artificial Analysis publishes the rates in its data API. LiteLLM publishes a JSON file of prices and context windows on GitHub. Search the exact model id and note the day you opened the file. A credit grant needs an amount, an expiry, and whether leftover credit vanishes.
Read the license and the token price
I may replace one paid chat task with an open model or a cheaper API. I will paste a model-card URL and a price-page URL. Using only those pages, list: the license name and a link to the license text; whether commercial use is allowed as written; any form, acceptable-use rule, or non-commercial clause; context length if stated; input and output price per million tokens, with currency, if stated. Mark anything missing as not stated. Then list which facts I still need before I drop a subscription: privacy, file uploads, tool use, and how it handles my language. Do not invent a price. Model card: [paste the model card] Price page: [paste the price page]
Expert
Rental rates, home listings, and dataset licenses
Date every card. On 20 October 2026, Runpod's pricing page, marked updated 8 October 2026, showed these rates on its Secure Cloud column. RTX 4090: 0.89 US dollars an hour, 24 GB VRAM (graphics memory). A100 80 GB PCIe and A100 80 GB SXM (two board formats): 1.79 US dollars an hour each. H100 PCIe 80 GB: 2.89 US dollars an hour. RTX 3090: 0.50 US dollars an hour, 24 GB VRAM, in that same column. It also lists lower Community Cloud rates, H100 PCIe at 1.99 US dollars an hour and A100 PCIe at 1.19. Container disk was 0.10 US dollars per GB a month while the pod runs. A volume disk costs 0.10 US dollars per GB a month while the pod runs and 0.20 when it is stopped, so a stopped GPU can still leave a disk bill. Read the live table again before you pay.
Vast.ai hosts set their own prices. On-demand runs until you stop it, billed per second. Interruptible is cheaper and can be taken back, so keep a checkpoint, a saved copy the job resumes from. Record the offer, rate, VRAM, and that flag.
There is no public source of prices for used graphics cards. The member market opens later. Until then, file a listing you opened that day: address, price, currency, VRAM, power if printed, and seller. Check a used RTX 3090 or an older card such as a Tesla P40 this way, and confirm current software still runs it before you mark it Runs at home. Add kronor only beside the exchange-rate page and its date.
Link the dataset card and quote the license id. From arXiv cs.AI recent, cite the paper id. From GitHub Trending, name the repository and the day.
Price one GPU job
Price one GPU job from the pages I paste. Use only those pages. Table columns: provider, offer name, GPU, VRAM, on-demand or interruptible, US dollars per hour, disk price and its unit, what happens when the job stops, and the date printed on the page. Add a total row for N hours of GPU time and D gigabytes stored for one month. Leave kronor blank unless I also paste an exchange-rate page; if I do, show the rate, the rate's date, and the kronor amount. Mark any missing figure as not stated. Flag a card that can be taken back mid-job, and say to keep a checkpoint. N hours: [number] D gigabytes: [number] GPU: [card name] Pages: [paste the price pages]
What every card records
Copy these lines onto each find.
- One-line name of the tool, model, credit, dataset, or listing.
- Who it is for: starting out, covering a paid chat, or running a job.
- The full https address of the page you opened.
- Checked at: calendar date, clock time, and timezone.
- Price, currency, and unit (per hour, per month, or per million tokens), or the page's own statement that it is free of charge.
- Whether a card, a phone number, or an account is required to start: yes, no, or not stated.
- What ends it: an expiry, a credit cap, an idle shutdown, or a rental the host can take back.
- The license id, when the find is a model or a dataset.
- Commission: none, or the program's name.
If the page gives no end date, write that, and open it again within seven days. The Runpod figures were read on 20 October 2026, and they go stale.
Commission and placement
If you are paid when someone follows your link, the first line of the card says so, names the program, and gives the offer's own address with the tracking code removed. When the archive opens, a card that hides a commission comes off the list. Use the same rule for anything you share before that. A checked free tier or a checked price can be a find. A shop cannot pay to be on the list.
The archive, when it opens
The tip form and the public archive open later. The filters planned for that archive are Free, Runs at home, and Price per million tokens. A card there will keep each checklist line, a review time, and a date to look again. Shop prices for used hardware will stay tied to a listing a person opened that day. There is no open feed of those shop prices. The market and the job board have not opened yet. The forum is open to invited members during the trial.
If Polaris buys the shared station, jobs that need more graphics memory than a home computer has could run there. About the shared station
Sources
- Google Colab FAQColab's own words on free-of-charge use, restricted free GPUs, unpublished limits, and the 12-hour cap for a free notebook.
- Practical Deep Learning for CodersThe free course. Phone verification for a Kaggle GPU is stated on the course Kaggle page.
- Hugging Face models, Apache-2.0 filterThe public model list with the license filter. Use the MIT filter the same way, then open the model card and the LICENSE file.
- Arena leaderboard datasetPublished preference scores with a license string. The agent latest split read here was dated 19 October 2026. License text still has to be checked on the model card.
- Artificial Analysis Data APIDocumented API for model prices per million tokens, benchmarks, and speed. Free and paid access tiers are described on that page.
- LiteLLM model prices and context windowsA large public JSON file of model prices and context windows in the LiteLLM repository. Search for the exact model id and note the day you opened it.
- Runpod GPU cloud pricingHourly GPU rates and disk prices. The page was marked updated 8 October 2026 and was read on 20 October 2026 for the hourly rates quoted here.
- Vast.ai GPU pricingLive marketplace rates set by hosts, with on-demand, interruptible, and reserved described on the page. Copy the offer you actually see.