07 — Choosing a Model Provider: Claude, GPT, Gemini, Cohere, Mistral, and Hugging Face

My default answer to "which model?" used to be "just call OpenAI," and it stopped working the moment a real choice came up. The frame that made the landscape readable: choosing a model is a multi-axis decision — capability, cost, context length, modality — and the provider is a separate decision from the model. [1] Hugging Face isn't a model, it's a hub; Claude and GPT are both models and providers; "use Gemini" conflates the model with the company that hosts it. Untangling those collapses most of the confusion.

The framing that finally landed is a single matrix I now fill in per feature. Down the rows: the candidate models. Across the columns: the axes that actually matter — raw capability on my task, price per million tokens, context window size, which modalities it handles, and whether I'm allowed to self-host. The model that wins on capability isn't always the model that wins the feature, because cost or latency or data sensitivity can disqualify it. The provider decision is then "who runs this model for me," and that's where Hugging Face comes in as the open-weights clearing house.

The closed-model providers

The big four closed providers each have a distinct center of gravity [2][3][4][5].

Anthropic's Claude is positioned around safety and alignment — Claude is named after Claude Shannon, the father of information theory, and the model is built to minimize harmful outputs and stay aligned with human intent [2]. In practice, Claude has earned a reputation for careful instruction-following, long-form reasoning, and strong coding performance. I reach for Claude when the task rewards careful adherence to a complex system prompt or when I want structured, low-hallucination output.

OpenAI's GPT and o-series is the broad default. The GPT family excels at natural language tasks — generation, translation, question answering — and the o-series covers specialized models like Whisper for speech [3]. OpenAI's platform is the most mature ecosystem, with the most documented patterns, which makes it the easiest first call when I'm prototyping.

Google's Gemini is multimodal-first. It was designed from the start to integrate text with images, video, and audio, and it pairs generative capability with reasoning skills [4]. Gemini is the natural choice when the feature genuinely needs to reason across modalities — analyzing a video, reading a chart — rather than text alone.

Cohere is enterprise-focused. Its models are tuned for production NLP — search, classification, retrieval — with an emphasis on safety, reliability, and ease of integration [5]. Cohere is the provider I look at when the use case is enterprise search or retrieval-heavy and I want a vendor whose whole pitch is "production NLP," not general chatbot.

The open-weight provider: Mistral

Mistral AI sits on the open side. It develops open-weight models designed to be efficient and versatile, releasing the trained weights so developers can customize and deploy them [6]. The strategic value of Mistral (and the open models generally) is that I get a capable model _and_ the option to run it myself, fine-tune it, or move it between providers. The tradeoff is that I take on more of the operational burden if I self-host.

Hugging Face: the hub, not a model

The piece I had to disentangle is Hugging Face. It's not a single model or even primarily a model provider — it's the leading open platform for AI, best known for its Transformers library and the Hub, a central repository where anyone can publish and download pre-trained models, datasets, and demos [7]. When I "use a model from Hugging Face," I'm pulling a model someone trained (often a company like Meta, Google, or Mistral) from a shared registry. The Hub provides versioning, model cards, and deployment tooling on top.

The distinction matters because it changes how I think about supply. The closed providers are single sources — I get the model from the company that made it. Hugging Face is a market — I pick from thousands of models by many creators, and the quality and licensing vary wildly. The Transformers library is the common API that lets me use most of them with consistent code, which is what makes the ecosystem navigable [7].

The decision axes, concretely

The matrix I fill in has these columns:

• Capability on the task. Not abstract benchmark scores — performance on _my_ eval set. A model that wins on a leaderboard can still lose on my specific data.
• Cost per token. Both input and output. For high-traffic features this dominates; for prototypes it's barely relevant.
• Context window. How much the model can consider at once. Long contexts help RAG and long conversations but cost more.
• Modalities. Text-only, or text plus image/audio/video. Multimodal models cost more but unlock features text-only models literally cannot do.
• Openness and hosting. Can I self-host if I need to? Is the data allowed to leave my network?

The model that's best on every axis doesn't exist. The point of the matrix is to make the tradeoff explicit so I pick deliberately rather than by default.

How I use this

The discipline is to stop reaching for one provider out of habit. For a prototype I'll grab whichever closed model is easiest to call — usually OpenAI or Claude — because ecosystem maturity speeds me up. For a feature with a tight cost ceiling I'll benchmark a smaller open model from Hugging Face against the default and often find it's good enough at a fraction of the price. For multimodal features I start with Gemini. For enterprise search I look at Cohere. And I keep the provider decision separate from the model decision: a model available from multiple providers lets me shop on price and latency, which is why OpenAI-compatible APIs and routers (covered later) matter. The matrix is what turns "I just use X" into a real engineering choice.

References

[1] BentoML, "Choosing the right model," 2024. [Online]. Available: https://bentoml.com/llm/getting-started/choosing-the-right-model

[2] Anthropic, "Claude 101," 2024. [Online]. Available: https://anthropic.skilljar.com/claude-101

[3] OpenAI, "OpenAI Models," 2024. [Online]. Available: https://developers.openai.com/api/docs/models

[4] Google, "Google AI Training," 2024. [Online]. Available: https://grow.google/ai/

[5] Cohere, "LLM University," 2024. [Online]. Available: https://cohere.com/llmu

[6] Mistral AI, 2024. [Online]. Available: https://mistral.ai/

[7] Hugging Face, "Hugging Face Documentation," 2024. [Online]. Available: https://huggingface.co/docs/hub/en/index