Free API Key Hunting (Mid-August Update)

July 09 - 2026

Mid-August Update: Google updated its pricing and thus Gemini Flash (non Lite) is available within the free tier! The API access needed for JASP AI integration works for Flash 3.5, 3.6, and 3.7 – even if 3.7 is not yet available in the Gemini App or Browser Chat in Europe. To use it switch to Gemini Flash this way: JASP => Main Menu => Preferences => AI => Provider: Google Gemini => Model => Gemini-3.X-Flash – where you change X to your preferred number. Currently the higher version numbers give you faster response times (3.5 to 3.6) and more intelligence (3.6 to 3.7) at the cost of more hallucinations (all 3 Flash versions around 60%) compared to Flash-Lite 3.5 at 34%. For more info click the big table below to get to the updated ranking.

End of July Update: Google updated Gemini Flash-Lite to version 3.5 (21. July 2026). To use it: JASP => Main Menu => Preferences => AI => Provider: Google Gemini => Model => Gemini-3.1-flash-lite=> Change the 1 to a 5 => Gemini-3.1-flash-lite. This is a huge upgrade especially in regard to hallucinations (down from 83% to 34%).

JASPs AI agent needs an API key. The application programming interface allows JASP to communicate with any AI provider, be it Google, OpenAI, Anthropic, or else. It is needed because current Laptop and Desktop computers mostly are not fast enough. They could run local AI agents offline, but their “intelligence” would pale in comparison to the capabilities of online ChatBots we are used to. Agent means that you cannot only chat with the bot, but that it can also click for you. Tell it, which analysis goal you have and it will produce graphs, tables and annotations using JASP. The clicks are of course restricted to the JASP interface, so the bot cannot go rogue and delete your files or worse.

The Trouble

Most providers want you to throw your wallet at them and pay per token. A token is roughly equivalent to a syllable. And you pay not only for the output tokens of the AI (model), but also for the input (your prompt aka question or task) and the “thinking” tokens needed by the model in the background. Fortunately there are cheap and free models. And when the cost trend of the past 4 years continues, we will see more capable free tier models go online every few months.

The Sources

There are many models available to choose from in JASP. You can even configure a custom provider with a custom model. So we need data to show us which model is the best (free) one for our use case. The most comprehensive and up to date sources for this are two information aggregators and benchmark creators: EPOCH (https://epoch.ai/benchmarks) and Artificial-Analysis (https://artificialanalysis.ai). Both send the bots through a number of hard to solve parcours (benchmarks) across all kinds of domains (math, science, law, economy, coding, agentic tasks, and so on). Both are independent from the AI labs considering their business model and transparency so far. To decide on a model, I am mostly using Artificial-Analysis. They present the current situation in a more accessible way and update a bit more often. Their website is packed with bar charts, scatterplots and time series – a data analyst’s dream come true.

The Criteria

There are criteria we as data analysts might want to look out for:

AA Intelligence / Language capabilities: The “Artificial Analysis Intelligence Index” is currently comprised of 9 evaluations (GDPval-AA v2, 𝜏³-Banking, Terminal-Bench v2.1, SciCode, Humanity’s Last Exam, GPQA Diamond, CritPt, AA-Omniscience, AA-LCR). All models run through all 9 benchmarks. The resulting score is a weighted percentage of correctly solved tasks, championed by Anthropics Claude Fable 5 at the moment (60%). This benchmark evaluates English text only.
AA-Omniscience / Low hallucination rate: In science we probably want a low hallucination rate, which is measured by “AA-Omniscience” already included within the 9 benchmarks. But maybe you want its weight higher than the current 12%. Omniscience rewards correct answers, penalizes hallucinations, and has no penalty for refusing to answer. Scores range from -100 to 100, where 0 means as many correct as incorrect answers.
MMMU Pro / Vision capabilities: For statistical data analysis we need graph / plot analysis too (look out for the flag “image input”). Consequently the “MMMU Pro” vision benchmark is based on 1.7k questions which require interpreting and reasoning over images.
SimpleBench / Common sense and imagination: SimpleBench and its sibling LLMcouncil cover spatio-temporal reasoning, social intelligence, and trick questions with 200 multiple-choice questions. This benchmark is covered by Epoch.

The Data

Gathering the numbers there appeared some missings (grey), which were imputed by interpolating from results of earlier model versions. To get to a final score I transformed the -100 to 100 range of the AA-Omniscience bench back to percentage scale and took the average of the benchmarks. Click for a constantly updated version of the table.

Three Hubs

Free API access is available with Google’s Gemini 3.X Flash (Lite). Of course “for free” means that you pay with your data and that there are some per hour and per day limits. It sports very good availability since Google is ranked within the Top 5 of the largest data center operators in the world, dwarfing most AI labs currently in the race. Visit this tutorial on how to set up Flash Lite.

Then there are the hubs of OpenRouter and Nvidia. Both offer free but also (more heavily) rate limited API keys through their platforms. Nvidia’s limit: up to 40 requests/minute – whatever that means. OpenRouter tells us: “If you have purchased at least 10 credits, your free model rate limit will be 1000 requests per day. Otherwise, you will be rate limited to 50 free model API requests per day”, that surely will not last long.

OpenRouter currently lists 19 free models with at least a 128k context length. You can try models even if they do not show the “Image” input modality, because most models reroute images to seperate vision models like DeepSeek V4. It started without vision completely and got it later bolted on due to community pressure. Nvidia lists 57 free models but one cannot filter for context length. On both platforms during rush hours you may have bad luck and will catch a “time out”, “not available”, or “status code 40x” errors, also depending on the popularity of the model you chose. From my experience bots routed through Nvidia’s platform have a higher availability and answering speed than through OpenRouter and Google offers even more speed and availability.

Free Open Weight Models

Since Googles Gemini is completely closed source and weights you might prefer open weight models like MiniMax, GLM 5.x, or Qwen3.x. The solution is indeed an account at build.nvidia.com or openrouter.com. After registering create an API key at https://build.nvidia.com/settings/api-keys or https://openrouter.ai/workspaces/default/keys. Then go to JASP settings => AI, activate it, and choose from the “Provider” drop down:

“Custom Provider” for Nvidia with “Endpoint URL” https://integrate.api.nvidia.com/v1/chat/completions or
OpenRouter with “Model: Custom” (Endpoint URL is automatically filled in)

Copy paste your API key into the empty “API key” field. Now choose a model and put it into the “Model” field. If the model is on OpenRouter and you want to try that, you need to append “:free”. Finally hit “Try connection”: