Free LLMs have gotten genuinely good. The ceiling on "no credit card, no API key" has risen dramatically since 2024; several free tiers now run models that would have been frontier-class 18 months ago. But each one has a specific breaking point, and they're not all the same. This is a map of eight options, what they're actually good for, and where each stops working.
The three consumer chatbots
These run in a browser. No setup, no billing, no code.
ChatGPT Free now runs GPT-5.5 Instant, the same model OpenAI set as the default for all users on May 5. The practical limit is roughly 15-40 messages per 3-hour window; OpenAI doesn't publish the exact number, and it shifts with server load. When you hit the ceiling, or during peak hours, the interface silently falls back to GPT-4o mini. Web search and basic DALL-E 3 image generation (2-3 per day) are included. The wall: no API access, no memory on the free tier, no custom system prompts.
Claude Free gives you Claude Sonnet 4.5 at about 15-40 messages per 5-hour window. The quality ceiling here is high. Sonnet 4.5 handles long documents, structured reasoning, and code review well within the context limit. Projects, which let you maintain a persistent knowledge base across sessions, are restricted to paid plans. The wall: no API, session history resets, no Projects.
Gemini Free runs Gemini 2.5 Flash with limited access to 2.5 Pro. The standout feature is Google ecosystem integration: it reads Gmail, Docs, and Drive, and is connected to real-time web search. That integration makes the free tier genuinely more capable for research tasks than the chat-only options above. The wall: since April 2026, Gemini 2.5 Pro is fully behind the Gemini Advanced paywall ($20/mo). Deep Research is restricted. If you want 2.5 Pro quality in a chat interface, you're paying.
DeepSeek: the free tier that punches hardest
chat.deepseek.com runs DeepSeek V4 Pro with no daily message cap. That's the same model that in our April benchmarks beat GPT-5.5 on reasoning and coding tasks at roughly one-fifth the API cost. For raw chat volume on drafting, analysis, or working through a hard problem, nothing on the free list comes close.
The caveats are real and worth naming directly. DeepSeek is a Chinese company with China-hosted servers. Its data retention policy isn't equivalent to US or EU alternatives. For personal projects, exploring ideas, or learning, that's probably fine. For client work, anything under NDA, or anything with personal data in the prompt, it isn't.
Free API access for builders
Once you're writing code that calls a model more than once, consumer chat interfaces stop working. These three give you an actual API endpoint without a credit card.
Google AI Studio offers a permanent free tier for the Gemini API: Gemini 2.5 Pro at 5 requests/minute and 100 requests/day, Gemini 2.5 Flash at 15 requests/minute and 1,500 requests/day. For a prototype or overnight evaluation run, 1,500 Flash requests per day is workable. The wall: 100 Pro requests/day stops any batch job in its tracks. Prompts sent via AI Studio are used for model training by default. Toggle that off in settings before you paste anything you care about.
Groq's free API runs Llama 3.3 70B and Llama 4.1 70B at hardware-accelerated speeds. Community benchmarks consistently show 300+ tokens/second, which makes it the fastest free-tier option by a significant margin. No credit card is required, and the model catalog is stable. The limits are 30 requests/minute and 6,000 tokens/minute per model. That's enough for prototyping; it's not enough for any user-facing feature with meaningful traffic.
OpenRouter free models include 29 permanently free options as of May 2026: DeepSeek V3, Llama 3.3 70B, Qwen3-Coder, Gemma 3 27B, and more. The interface is OpenAI-compatible, which means switching your existing app takes one line of code. The wall: free models share capacity, and latency can spike 5-10x during busy windows. Build in a timeout guard before you put this in front of a user.
How they compare
| Provider | Model | Access | Rate limit | Privacy note | Upgrade cost |
|---|---|---|---|---|---|
| ChatGPT | GPT-5.5 Instant | Chat | ~30 msg/3h | OpenAI ToS | $20/mo Plus |
| Claude | Sonnet 4.5 | Chat | ~30 msg/5h | Anthropic ToS | $20/mo Pro |
| Gemini | 2.5 Flash / Pro | Chat + API | 1,500 RPD (Flash) | Google ToS | $20/mo Advanced |
| DeepSeek | V4 Pro | Chat | No cap | China-hosted | API pay-per-use |
| Groq | Llama 3.3/4.1 70B | API | 6K TPM | US-hosted | Pay-per-use |
| OpenRouter | V3, 70B, Qwen+ | API | Shared capacity | Varies | Pay-per-use |
When free stops being enough
Three signals that the free tier is genuinely your constraint, not just a temporary annoyance:
You're hitting message caps before noon. If 30 messages over 3-5 hours isn't enough for a working session, you're past the prototype phase. The cap isn't an edge case at that point.
You need a loop. Any code that calls a model more than a handful of times needs an API key. Consumer chat interfaces don't expose one. Google AI Studio's free tier is the exception, but its per-day limits become the constraint quickly.
Your users are waiting on it. Shared free-tier capacity means unpredictable latency spikes. A p95 of 30 seconds from OpenRouter free isn't an implementation bug. It's the business model. If latency is a product decision, free shared capacity isn't a foundation.
When you're ready to move off the free tier, the question shifts to which paid model fits the task, not which is biggest. How to choose an LLM in 2026 covers the five axes that actually matter for that decision. If you want to test multiple providers before committing, LLMTest routes across all of them from a single endpoint with per-call cost tracking. Our benchmark methodology explains how the comparisons work if you want to run your own before picking.