Skip to main content

AI providers · Dracon Research Team

AI free tiers in 2026: the limit, the card question, and the date we checked

Every week the same question arrives in twenty small variations: is there a free key for this model, and if so, what does it actually let you do. The key-setup pages answer the first half for each provider. This page answers the second half for all of them at once: the limit, whether the tier asks for a card, and the date we last read the provider's own page. Four names readers expect to find here are missing on purpose, and the second table says why.

The board: every free tier we track

Meaningful recurring tiers sort first, then trial and funnel tiers. Nine rows carry no hard number because the provider publishes none in a stable form: the quota rotates, varies per model, or is qualitative, so the row says that instead of inventing a figure. The card column answers one question: does this free tier ask for a payment card. None of them does: every row on the board, Codex included, is usable without putting a card on file.

Free-tier limits, per provider (verification window 2026-10-26 to 2026-10-27). Meaningful recurring tiers first, then trial and funnel tiers.
ProviderFree-tier limitCardVerified
ClineExtension free for individuals; inference usage-based or BYOK (limits vary per model).No card2026-10-27
Cloudflare Workers AI10K neurons/day free; overage $0.011 per 1K neurons.No card2026-10-27
Google AI StudioFree tier with free input/output tokens on select models; no billing account required.No card2026-10-27
GroqFree Plan limits per model: up to 30 RPM and 1K-14.4K req/day (docs rate-limits table).No card2026-10-27
Kilo CodeFree for individuals with sign-in; per-model request limits rotate.No card2026-10-27
LLM Gateway20 req/min on free modelsNo card2026-10-27
Mistral Le ChatBase-tier messages, web searches, libraries, and image generation (subject to fair use; check Mistral's pricing page for current specifics)No card2026-10-27
NVIDIA NIM~40 RPM, no daily capNo card2026-10-26
OllamaLocal inference free and unlimited; cloud pay-as-you-go at $0 with 3 concurrent requests.No card2026-10-27
OpenAI CodexCodex is included on ChatGPT Free (no card; limited access) and on paid plans from Go up; Codex Cloud needs a paid tier. Usage limits vary by plan: check Settings or the usage dashboard, or enter /status in an active Codex CLI session.No card2026-10-27
OpenCodeNine limited-time free models via Zen (model list rotates).No card2026-10-27
OpenRouter50 req/day on the base free tier (no card); $10 in PAYG credits unlocks 1,000 req/day of free models (free model pool rotates; check the provider's pricing page)No card2026-10-27
OVHcloud AI EndpointsFree trial of every model via the AI Endpoints catalogue (no key); keyless use is rate-limited.No card2026-10-27
SiliconFlowFree models at fixed rate limits (chat models about 100 RPM) after real-name verification; usage billed ¥0.No card2026-10-27
Alibaba Coding PlanAutomatic free quota on eligible Singapore-region models for new users (product page markets the trial as 10M tokens).No card2026-10-27
BytePlus ModelArkNew-user free trial quota via the ModelArk console (amounts rotate).No card2026-10-27
Mimo50 one-time free credits for the Airo site builder.No card2026-10-27
Z.AI GLMZ.ai chat is available free with current GLM access; the API and coding plans are paid with separate limits.No card2026-10-26

Four providers that look free and are not

Each of these had a free story in the past: a credit that expired, a beta that ended, a service that retired. What their pages say now is below, with the date we read them. If any of them launches a real tier, the generator behind this page refuses to publish until the verdict is rewritten.

Providers checked and found to have no free tier (verification window 2026-10-26 to 2026-10-27).
ProviderWhat the page actually saysChecked
DeepSeekPaid-only per-token API pricing; no free tier, trial or new-account credit advertised (pricing page, checked 2026-10-26 and 2026-10-27).2026-10-27
GitHub ModelsRetired 2026-07-30: the playground, model catalog, inference API and BYOK are unavailable to all customers (GitHub Docs, checked 2026-10-27).2026-10-27
Hugging FaceFree accounts receive no monthly Inference Providers credit and must purchase credits; monthly credits go to PRO ($2) and Team/Enterprise seats only (provider docs, checked 2026-10-27).2026-10-27
HyperbolicInference requires a Pro plan (one-time deposit of $5 or more); personal accounts on the Free Tier cannot use inference (billing docs, checked 2026-10-27).2026-10-27

How to read a limit

Requests per minute and per day are the two dials that matter; neurons are Cloudflare's unit of inference work, and ten thousand of them a day is enough to build and test against. A daily quota that renews beats a one-time credit by an order of magnitude, which is why the board sorts recurring tiers first and why one-time trials sit at the bottom with the funnel tiers. The verification window on each table is the oldest and newest provider check behind its rows.

How we checked, and what would change our answer

On 2026-10-26 and 2026-10-27 we re-read every provider page behind these rows: static pricing pages and docs directly, JavaScript pages through a rendered reader, and the retirement notice for GitHub Models in the provider's own docs. Quotas the page no longer states were cut from the row rather than carried forward, and four rows were flipped to no-tier with the reason recorded. A drift test fails the site build when the catalog moves, so this page is regenerated instead of rotting: a new limit, a new card requirement, or a resurrected tier breaks the build until the row is rewritten.

Sources & provenance