Why is it free?
Free tiers: a busy GPU makes tokens for cents, and free buys customers, data or idle hours · Beginner · 5 min · llm · after GPUs: thousands of cores and fast memory for matrix multiplies, Model size, free vs paid tiers
Every answer costs GPU time, rented at about $3–4 an hour. A busy GPU writes a million tokens for cents, so free tiers are cheap to run: a bet on a future customer, your data or idle GPU hours.
1 Guess first
An answer runs on a GPU that is rented or owned by the hour, busy or not. It serves many users at once, so the cost of one answer is hourly price ÷ tokens written per hour. The busier the GPU, the cheaper each token.
3 Look
What one answer costs, and what free buys
1 / 5Rent one GPU, on demand, per hour RunPod Community Cloud H100 SXM $2.69 RunPod Secure Cloud H100 SXM $3.99 Lambda, 8-GPU instance H100 SXM $3.99 Lambda, 1 GPU H100 SXM $4.29 Vendors’ own pages, read 2026-10-26; prices change
Every answer runs on a GPU someone rents or owns. A top data-center GPU costs about $2.70 to $4.30 an hour on demand.
Free offers we track: free API credits · startup credits · when each was last checked: freshness
On a busy GPU a million tokens cost cents; an idle one bills the same hour for nothing. Free tiers fill that gap and are priced to pay back.
4 You are the model
You run pricing at an AI company. For five real free deals, say what each one buys you: a future customer, data, idle GPUs, or a showcase.
deal 1 of 5 · what does it buy the company?
Gemini API free tier: free tokens, and human reviewers may read your prompts. On the paid tier they aren’t used.
Rules from the vendors’ pages, read 2026-10-26
You run pricing. Name what each free deal buys the company.5 Check
Because free is a bet, it gets re-priced. In aifree.dev’s own changelog for October 2026 (changes recorded October 13–24), 3 offers cut their stated amount: the Runware Startup Program from $30,000 to $1,000. 4 offers started asking for a card. Free means free until the numbers change.
7 Say it back
8 Mission · 2 min
Open the freshness table, pick an offer whose last check says “Change detected”, and open its page. Read what changed, and decide which of the four reasons the change fits.
Three quick ones
Drawn at random from this lesson's pool; different next time.
Done?
Get an email when new lessons come out
New lessons arrive in the weekly digest, only in a week that has one. You confirm by email first; unsubscribe in one click.
Go deeper
Sources
- DeepSeek, “DeepSeek-V3/R1 Inference System Overview” (open-infra-index, Open Source Week day 6; https://github.com/deepseek-ai/open-infra-index/blob/main/202502OpenSourceWeek/day_6_one_more_thing_deepseekV3R1_inference_system_overview.md, figures for 2025-02-27 12:00 to 2025-02-28 12:00 UTC+8, checked 2026-10-26): H800 rental assumed at $2 per GPU-hour, 8 GPUs per node; ~14.8k output tokens/s per decode node; daily GPU cost $87,072; 608B input and 168B output tokens; $562,027 a day if all tokens were billed at R1 prices ($0.14 / $0.55 input, $2.19 output per 1M), a 545% cost profit margin; actual revenue lower because V3 is cheaper, only a subset of services is monetized (web and app access remain free) and night discounts apply; nodes reduced at night and given to research and training. $0.30 per 1M output tokens is our arithmetic: $16 an hour ÷ (14,800 × 3,600 = 53 million tokens an hour)
- RunPod pricing (https://www.runpod.io/pricing, page updated 2026-10-08, checked 2026-10-26): H100 SXM $3.99/hr Secure Cloud, $2.69/hr Community Cloud
- Lambda pricing (https://lambda.ai/pricing, checked 2026-10-26): H100 SXM $4.29 per GPU-hour for 1 GPU, $3.99 per GPU-hour on the 8-GPU instance
- DeepSeek API, Models & Pricing (https://api-docs.deepseek.com/quick_start/pricing, checked 2026-10-26): deepseek-flash (DeepSeek-V4.1-Flash-1017) output $1.2 per 1M tokens at peak, $0.6 off-peak; “Off-peak rates are half of the peak rates”; peak 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday
- Google, Gemini Developer API pricing (https://ai.google.dev/gemini-api/docs/pricing, last updated 2026-10-22, checked 2026-10-26): gemini-2.5-flash-lite free tier “Free of charge”, paid tier $0.10 input / $0.40 output per 1M tokens; “Used to improve our products”: Yes on the free tier, No on the paid tier. Gemini API Additional Terms (https://ai.google.dev/gemini-api/terms, last updated 2026-04-28, checked 2026-10-26): on Unpaid Services “human reviewers may read, annotate, and process your API input and output”; Paid Services: “Google doesn’t use your prompts or responses to improve our products”. Rate limits (https://ai.google.dev/gemini-api/docs/rate-limits, last updated 2026-10-15, checked 2026-10-26): Tier 1 requires “Set up and link an active billing account”
- OpenRouter, API rate limits (https://openrouter.ai/docs/api-reference/limits, checked 2026-10-26): :free model variants 20 requests per minute; 50 requests per day with less than 10 credits purchased, 1,000 with at least 10
- aifree.dev offer changelog (offer_events, counted 2026-10-26 for changes recorded 2026-10-13 to 2026-10-24): 3 stated amounts cut (Runware Startup Program $30,000 to $1,000, AgentRouter $200 to $25, CleanAPIs 5,000,000 to 500,000 tokens a month), 4 offers newly requiring a card
- The 50%, 25% and 10% loads in the viz are illustrative: the same node and hourly price with fewer tokens per second
Published Oct 26, 2026 · updated Oct 26, 2026