Why You Should Use Open Source Models

A running list of reasons.

Last updated 22 June 2026

  1. So you get better value for money

    Closed-source model providers make healthy margins on their API pricing, and they need to pay back the high training costs. For the same capability level, open models are 5–10x cheaper.

    Sources: Doubleword — comparison pricing page · Artificial Analysis — model comparison

  2. So you can get frontier performance

    Open-source models are now highly competitive. For example, as of 22nd June 2026, GLM 5.2 is reported to be beating GPT-5.5 on design benchmarks.

    Sources: Doubleword — frontier open-source LLMs · MindStudio — what is GLM 5.2

  3. So you control depreciation / upgrade cycles

    OpenAI / Anthropic regularly deprecate models that are no longer their “frontier” models. E.g. GPT-5 is scheduled to be deprecated on December 11, 2026.

    Source: OpenAI — model deprecations

  4. So you don’t lose access

    US-based access to frontier models can be removed within 90 minutes, as evidenced by the June 10th Fable 5 event.

    Source: Anthropic — Fable Mythos access

  5. So you don’t give money to a potential future competitor

    Frontier model providers may build businesses in the future that compete with yours. As an example, if you were running a consulting business, frontier-model joint ventures are now a competitor. Or if you are Figma, Anthropic is now a competitor.

    Sources: TechCrunch — Anthropic CPO leaves Figma’s board · Forbes — OpenAI starts selling what your consulting business sells

  6. So you don’t discriminate your employees based on nationality

    US-based frontier models may discriminate access based on user nationality, forcing you to have tiered access within your organisation.

    Source: Anthropic — Fable Mythos access

  7. So you avoid a single nation dominating the intelligence business

    All of the competitive frontier model labs are US-based, and the US government has already demonstrated an appetite to use their businesses and industries as leverage in geopolitical conversations. The US government has also demonstrated that it is happy to restrict access to models on a nationality basis.

    Source: Anthropic — Fable Mythos access

  8. So you can control your latency, throughput, and uptime

    Proprietary models have inconsistent latency / throughput / uptime metrics that can’t be tuned for your use case. Open models can be deployed in a way that suits your use case by choosing a provider that optimises its stack in line with your requirements.

  9. So you get consistent model performance

    Proprietary model providers have been known to nerf model capabilities without warning — sometimes for ‘safety reasons’, sometimes accidentally, and sometimes for suspected capacity reasons. Open models don’t change their capabilities over time, because the weights don’t change. If you are building an application on top of a model, you know the model’s capabilities will not change over time.

  10. So you can improve performance on your use case

    You can further fine-tune / post-train open models with your proprietary data to make them better at your use case. Proprietary model providers have a built-in capability ceiling that you can’t cross.

  11. So you don’t get locked into a single provider

    Open-model inference providers have to compete aggressively on cost, reliability, data policies, locality, latency, and more. They compete aggressively to ensure they have the best serving stack and developer experience, because they can’t differentiate as well on the underlying model. Also, by building with open models you will adopt harnesses and tools that are model-agnostic by design — whereas proprietary model providers attempt to lock you in with a full harness ecosystem (e.g. Claude Code).

  12. So you don’t need to go through identity checks to access it

    As of June 2026, Claude now requires identity verification for some use cases.

    Source: Identity verification on Claude

  13. So you benefit from a fast-moving, vibrant community

    Open models get a whole ecosystem on top of them: quantizations, fine-tunes, and serving tooling (vLLM, SGLang, Dynamo, OpenCode), built and improved by thousands of people, often within hours of release.

  14. So you can host it in an airgapped environment

    If you need to deploy in an airgapped or highly secure environment, only open models will allow you to do this (unless you’re the US government). You can take an open model and deploy it wherever you have access to compute, without any need for a connection to the internet.

  15. So you can run on the edge / on-device

    Open weights can be run directly on laptops, phones, robots, or embedded hardware when use cases require.

  16. So you can control your carbon footprint

    Open models allow you to control where you deploy and what energy source powers that data centre.

  17. So you can achieve data sovereignty

    Open models can be deployed wherever you need them to be, and can ensure that data does not leave your required data region. Proprietary models are only deployed in a selection of data regions. For example, if I am a highly regulated Mexican company or government, I cannot select Mexico as a data region in Azure or AWS Bedrock — meaning, in order to get access to these models, I need to send my data externally.

  18. So you don’t have to worry about others training on your data

    Open-source inference providers build their business on top of providing the best inference experience for their customers — they are not in the business of training the next generation of generalist models. Therefore, open inference providers have no incentive to do anything funny with your data or train on it without your explicit consent. For proprietary model providers, there are legal guarantees, but the incentives are not as aligned.

  19. So you can distil into smaller task-specific models

    With open weights, you’re free to distil a large model down into a smaller, cheaper, task-specific model — one that runs faster and costs far less to serve, while keeping the capability you actually need. DeepSeek, for example, distilled its R1 reasoning model into 1.5B–70B Qwen and Llama checkpoints, with the 32B distil outperforming OpenAI’s o1-mini on several benchmarks. You can’t legally do this with closed models: OpenAI, Anthropic, Mistral, and xAI all prohibit using their outputs to train competing models in their terms of service.

    Sources: DeepSeek-R1-Distill-Qwen-32B (Hugging Face) · Berkeley Law — AI distillation & closed-model terms