Free AI models: what you can actually use for free
Five consumer apps and four developer platforms offer permanent $0 tiers. Compare models, rate limits, data-use rules, and paywall boundaries.
Before paying for an AI plan, check what you can run without paying anything. Several major frontier labs and infrastructure providers maintain permanent $0 tiers. However, free offerings change constantly: models get swapped, rate limits fluctuate with cluster load, and usage policies differ sharply between consumer apps and developer APIs.
Last checked: 21 November 2026
Summary: what is free today
| Provider | What is free | Limits | App or API | Data-use note | Source |
|---|---|---|---|---|---|
| ChatGPT Free | GPT-6 Luna, mini / standard fallback | Dynamic prompt window based on traffic | Consumer Web / App | Prompts used to train models unless opted out in Settings | OpenAI Pricing |
| Claude Free | Claude 3.5 Sonnet / Haiku | Dynamic message limit resetting every 5 hours | Consumer Web / App | Not used for model training by default | Claude Limits |
| Gemini Free | Gemini 2.0 Flash / 1.5 Flash (Hydrogen) | Standard compute allowance resetting every 5 hours; 32k context | Consumer Web / App | Human review and model training active unless Activity paused | Gemini Limits |
| Grok Free | Grok 2 / Grok 3 mini | Rate-limited prompt pool (e.g. ~10–20 queries per 2 hours) | Web (X / grok.com) | Used for model training unless opted out in X settings | xAI Pricing |
| Perplexity Standard | Standard search engine routing | Unlimited standard searches; ~3–5 Pro searches every 4 hours | Consumer Web / App | Prompts used for training unless AI Data Retention toggle is disabled | Perplexity Pro |
| Google AI Studio | Gemini 2.0 Flash, 1.5 Flash / Pro API | 15 RPM / 1M TPM / 1,500 RPD (Flash free tier) | Developer API | Prompts & responses logged and used for Google model training | Google AI Studio Pricing |
| OpenRouter :free | Open-weight models (Llama 3.3, Qwen, DeepSeek) | ~20 RPM rate limit; subject to shared community capacity | Developer API | OpenRouter does not train, but upstream hosts have varying policies | OpenRouter Models |
| Groq Free Tier | Llama 3.3 70B, Llama 3.1 8B, Mixtral | Model-specific RPM/TPM caps (e.g. 30 RPM, 6,000 TPM) | Developer API | No training on customer prompts or completions | Groq Rate Limits |
| Cloudflare Workers AI | Curated catalog (Llama 3.1, Mistral, BGE, Flux) | 10,000 Neurons per day, resets at 00:00 UTC | Developer API | Cloudflare does not use customer inputs/outputs for training | Cloudflare Workers AI |
Every row was checked against the provider’s official page on 21 November 2026.
Free consumer tiers
Consumer chat applications provide a zero-cost entry point, but they meter access dynamically to protect paying subscribers during peak hours.
ChatGPT Free
- Models: Runs lighter-weight default models such as GPT-6 Luna or mini, with strictly metered, opportunistic access to higher-tier models during off-peak periods.
- Caps: OpenAI does not publish fixed prompt numbers for the free tier. Limits scale dynamically with server congestion; once reached, access falls back to lower-tier models or is paused until the rolling window refreshes.
- Paywalled: Unlimited GPT-6.1 Sol / GPT-6.2 access, Codex integration, ChatGPT Work, priority peak access, extended context windows, and advanced agent capabilities.
- Cheapest paid step: ChatGPT Go at $8/month (or ChatGPT Plus at $20/month). See ChatGPT pricing.
Claude Free
- Models: Claude 3.5 Sonnet and Haiku variants depending on server load.
- Caps: Anthropic meters usage through a dynamic 5-hour rolling limit. Heavy conversations or long attachments exhaust the session budget in fewer turns. When the threshold is reached, Claude displays a countdown timer until the 5-hour window resets.
- Paywalled: Claude 3.5 Opus, extended reasoning depth, Claude Code command-line agent, and 5× usage multipliers.
- Cheapest paid step: Claude Pro at $20/month (or $17/month billed annually).
Gemini Free
- Models: Gemini 2.0 Flash and 1.5 Flash (Hydrogen-class entry models).
- Caps: Google meters Gemini consumer apps via compute allowances that refresh every 5 hours inside a weekly ceiling. The free tier is constrained to a 32,000-token context window.
- Paywalled: Gemini 4 Neon (Pro) and Carbon (Ultra), 1-million-token context windows, Deep Research, 2× to 20× compute multipliers, and bundled Google One cloud storage.
- Cheapest paid step: Google AI Plus at $4.99/month (or Google AI Pro at $19.99/month). Compare options in How many tokens a month do Gemini plans give you?.
Grok Free
- Models: Grok 2 and Grok 3 mini on web and mobile.
- Caps: Users with standard X or grok.com accounts receive a periodic query pool (roughly 10 to 20 queries per 2 hours during normal load).
- Paywalled: Full-scale Grok 4.x reasoning, DeepSearch depth, priority image generation, and Cursor integration.
- Cheapest paid step: The X Subscription Plus at $8/month.
Perplexity Standard
- Models: Default search indexing and lightweight model synthesis.
- Caps: Unlimited basic search queries, accompanied by a small allotment of Pro searches (~3 to 5 queries every 4 hours or daily).
- Paywalled: Unlimited Pro search using top frontier models (Claude 3.5 Sonnet, GPT-4o), larger document uploads, and API developer credits.
- Cheapest paid step: Perplexity Pro at $20/month.
Free API access
If you need programmatic access for scripts, automation, or local tools, several providers maintain ongoing free tiers.
Google AI Studio (Gemini API)
Google AI Studio offers a free developer tier without requiring an upfront credit card charge:
- Models: Gemini 2.0 Flash, Gemini 1.5 Flash, and Gemini 1.5 Pro.
- Rate limits: Up to 15 Requests Per Minute (RPM), 1,000,000 Tokens Per Minute (TPM), and 1,500 Requests Per Day (RPD) on Flash models.
- Data-use rule: Under the Google AI Studio Terms of Service, all prompts and responses in the free tier are logged and used by Google to train and tune machine learning models, and may be reviewed by human contractors. If your workload involves proprietary code or private personal data, you must switch to the paid (pay-as-you-go) tier where training is turned off.
OpenRouter :free models
OpenRouter aggregates open-weights models through community-subsidized endpoints denoted with the :free suffix:
- Models: Meta Llama 3.3 70B Instruct, Qwen 2.5, DeepSeek R1/V3 free endpoints, and Mistral models.
- Rate limits: Approximately 20 requests per minute with per-account burst limits, subject to cluster availability.
- Data-use rule: OpenRouter does not train on customer inputs. However, requests are fulfilled by third-party hosting partners whose upstream logging terms apply.
Groq Free Plan
Groq provides high-speed LPUs for open models with a free developer tier:
- Models: Llama 3.3 70B, Llama 3.1 8B, Mixtral 8x7B, and Whisper.
- Rate limits: Generous developer caps (e.g. 30 RPM and 6,000 TPM for Llama 3.3 70B; 30 RPM and 14,400 TPM for Llama 3.1 8B).
- Data-use rule: Groq’s developer terms state that customer prompts and outputs are not used to train models. Operational logs are retained only for security and abuse detection.
Cloudflare Workers AI
Cloudflare provides serverless inference directly on its edge network:
- Models: A curated catalog of open models including Llama 3.1, Mistral, BGE embedding models, and Flux image generation.
- Allocation: 10,000 Neurons per day on the standard free Workers plan, resetting daily at 00:00 UTC.
- Data-use rule: Cloudflare explicitly commits that customer inputs and outputs processed by Workers AI are not used for model training.
Other developer free tiers
- Mistral AI: Provides a free tier on La Plateforme with basic rate limits for evaluation and testing.
- Cerebras Inference: Offers a free developer tier with fast Llama 3.1 inference capped at 30 RPM.
- SambaNova Cloud: Provides developer access with rate-limited free API keys for Llama and DeepSeek endpoints.
Providers with NO ongoing free API tier
Be aware of popular misconceptions:
- OpenAI: Has no ongoing free API tier. Free trial credits expire, and ongoing API usage requires adding a payment method and pre-purchasing credits.
- Anthropic: Has no ongoing free API tier. Claude API access is strictly pay-as-you-go with prepaid credits.
- GitHub Models: Was retired on 2026-07-30. Do not rely on deprecated endpoints or older blog posts listing GitHub Models.
What “free” costs: data, throttling, and constraints
There is no free compute in AI infrastructure. When using a $0 tier, understand the structural trade-offs:
- Your data is often the payment: Consumer free tiers and Google AI Studio’s unpaid tier explicitly state that user inputs and completions are retained and reviewed to train future foundation models. Never send proprietary source code, confidential financials, or personal customer data through a free tier that trains on prompts.
- Dynamic throttling and degradation: Free tiers operate at lowest server priority. When clusters experience high demand, providers throttle free requests, lower rate limits, or downgrade models without prior notice.
- Truncated context windows: Free tiers almost always cut maximum context length (for example, 32k tokens on Gemini Free compared to 1,000,000 tokens on Gemini Pro).
Before spending money on a recurring plan, evaluate your actual volume. If your usage is light or irregular, these free tiers may be all you need. When your requirements exceed these limits, check Chat or coding: how to pick the first plan, compare buying a subscription vs paying for API tokens, or browse the official personal-plans catalog.
Last checked 2026-11-21 against the links below. Open the vendor page in your region before you pay — list prices move.
- Official · ChatGPT pricing & features · Open source · Checked 21 Nov 2026. Free tier runs GPT-6 Luna / mini with dynamic prompt limits; training opt-out in settings.
- Official · Claude usage limits & pricing · Open source · Checked 21 Nov 2026. Dynamic 5-hour reset window; no training on consumer chats by default.
- Official · Gemini Apps limits & feature availability · Open source · Checked 21 Nov 2026. 5-hour compute window, 32k context on Free. Unpaid prompts reviewed and trained unless activity disabled.
- Official · xAI pricing & Grok on X · Open source · Checked 21 Nov 2026. Free queries on X/grok.com rate-limited; training opt-out toggle in X account settings.
- Official · Perplexity Pro & Help Center · Open source · Checked 21 Nov 2026. Unlimited standard searches, limited Pro search queries.
- Official · Google AI Studio pricing & terms · Open source · Checked 21 Nov 2026. Free API tier up to 15 RPM / 1M TPM / 1,500 RPD on Flash. Prompts logged and used for training.
- Official · Groq console rate limits & terms · Open source · Checked 21 Nov 2026. Free developer plan with per-model RPM/TPM limits. No training on user inputs.
- Official · OpenRouter free models · Open source · Checked 21 Nov 2026. Endpoints with :free suffix; rate limits ~20 RPM, upstream host terms apply.
- Official · Cloudflare Workers AI pricing · Open source · Checked 21 Nov 2026. 10,000 Neurons/day free. Cloudflare does not train models on customer inputs.