🤝 How to access top AI models for fractions of a cent
Dear curious mind,
While open-source AI models are becoming a valid alternative to the closed models of the large US companies, running them locally can be complex and hardware-intensive. Today, we'll explore why using API access is the simplest and most cost-effective way to get started with these models, plus share practical tips on how to begin. Let's dive in!
In this issue:
💡 Shared Insight
API-keys: The Budget-Friendly Way to Use AI Models
📰 AI Update
Chinese Open-Source LLM MiniMax-01 Competes with Top AI Models
UC Berkeley Trained the Competitive Reasoning Model Sky-T1-32B-Preview for $450
🌟 Media Recommendation
Video: Zuckerberg on AI, AR, Privacy and Free Speech with Joe Rogan
💡 Shared Insight
API-keys: The Budget-Friendly Way to Use AI Models
Last week's issue explored why open-source AI models are needed alternatives to proprietary models from large US companies. While the most powerful open-source models like DeepSeek-v3 and Llama 3.3 70B are too large for consumer hardware, there's a cost-effective way to start using them: API access. Even if you plan to run smaller models locally for privacy reasons, testing them via API first lets you explore their capabilities without investing in hardware or waiting for slow responses.
What is API Access?
An API (Application Programming Interface) lets you use models on other people's servers. You receive an API key - think of it as your digital fingerprint - and send requests to their server. You pay only for your interactions, and if you're not a power user, this is a fraction of the monthly $20 ChatGPT subscription, making it perfect to experiment and find the best models for your needs.
Recommended Open-Source Models
There are far too many outstanding open-source models to name them all, especially as many are focusing on specific areas of expertise like coding or excelling in one or many non-English languages. However, the best general models, from my perspective, are in the following table:
Simplified Access through Aggregators
Most organizations behind the open-source models offer their own API service, including the currently best models from DeepSeek and the new competitor MiniMax (covered in the news section of this issue).
However, aggregation services like OpenRouter make it easier to start:
Single account to access various models
Competitive pricing (often matching direct API costs)
Low minimum balance of $5
Another option is Groq, an American company focusing on fast model execution speeds, currently offering free usage (no credit card required) with up to 30 requests per minute and access to the best Llama 3.3 70B model from Meta.
Comparing Chatbot Responses
You're likely using ChatGPT or one of its competitors for your AI chats. My recommendation is that you compare the outputs of one or multiple open-source models to the model you are currently using and judge if it's worth replacing. The easiest way to experiment with multiple models is through a Chrome extension called ChatHub, which can be used for free:
Chat with two models simultaneously
Use subscription services and / or API keys
Set up API access with a simple copy-paste of your key
If you like the tool, there is a life-time premium option, which you buy once to enable additional features:
Chat with up to six models simultaneously
Use local models via Ollama
Use the model aggregator OpenRouter
If you ever consider buying this option, it would help me if you use my ChatHub affiliate link. However, I recommend you to start for free and explore the tool without any costs.
Using the Chrome browser extension ChatHub allows you to experiment with multiple models at the same time and compare their outputs against industry leaders like GPT-4o.
Getting started is easy, and I encourage you to do so without any costs by using Llama 3.3 70B via Groq in ChatHub. I created a short video tutorial which you can follow:
If you are not happy with the results from Llama 3.3 70B, give the so far best open-source model DeepSeek-v3 a chance via the DeepSeek API which requires that you buy $2 in credits.
The next step would be to use OpenRouter which offers even more models via API. You can experiment with countless open-source models as well as nearly all models from big cloud players via API, some of them even entirely free. So even if you identify that you don't like the performance of the open-source models, you might find a valid alternative which is paid by usage compared to the fixed subscription costs of ChatGPT and alike.
This was part two of our three-part series on open-source AI models. Last week we explored the benefits of open-source models, and next week we'll dive into running models locally for maximum privacy.
📰 AI Update
Chinese Open-Source LLM MiniMax-01 Competes with Top AI Models [Model card on HuggingFace]
Following DeepSeek-v3's open-source release, MiniMax-01 emerges as China's second state-of-the-art language model to challenge leading global AI systems. Notably, MiniMax-01 features a 4-million token context window, doubling the previous record of 2 million tokens held by Google's Gemini Pro, marking a significant advancement in long-context processing capabilities. However, the current API-key only supports a context size of 1-million tokens, but this is still outstanding.
UC Berkeley Trained the Competitive Reasoning Model Sky-T1-32B-Preview for $450 [NovaSky Team article]
The model is built on Qwen2.5-32B-Instruct, shows competitive results in math reasoning and coding tasks compared to OpenAI's o1-preview. It's fully open-source, with all data, code, weights, and reports publicly shared.
🌟 Media Recommendation
Video: Zuckerberg on AI, AR, Privacy and Free Speech with Joe Rogan
Mark Zuckerberg, CEO of Meta, appeared on The Joe Rogan Experience for a 3-hour discussion covering Meta's latest AI initiatives, content moderation policies, and vision for the future of social media and Augmented Reality (AR) technology.
My key highlight from the conversation was his vision of using AI agents to replace mid-level engineers and write significant portions of code for their apps and services already in 2025.
My take: Replacing developers this year with AI is a bold statement, even if he adds that this will likely be expensive initially. However, recent user experiences with the hyped software developing agent Devon were disappointing [article: Thoughts On A month With Devin,𝕏 post by Ethan Mollick], but the AI space is moving forward at lightning speed and the results from coding agents might improve soon.
Disclaimer: This newsletter is written with the aid of AI. I use AI as an assistant to generate and optimize the text. However, the amount of AI used varies depending on the topic and the content. I always curate and edit the text myself to ensure quality and accuracy. The opinions and views expressed in this newsletter are my own and do not necessarily reflect those of the sources or the AI models.