Open-Weight vs Proprietary AI Models: Licences, Costs, Performance and How to Choose in 2026

How open-weight models such as Llama, DeepSeek, Qwen, Mistral and gpt-oss compare with proprietary APIs: what the OSI's open source AI definition requires, how licences differ, how big the performance gap is, the real costs of self-hosting and a framework for choosing.

PUBLISHED // UPDATED
DURATION // 06 MIN READ
Open-Weight vs Proprietary AI Models: Licences, Costs, Performance and How to Choose in 2026
FIG.01 // VISUAL DOCUMENTATION Open-Weight vs Proprietary AI Models: Licences, Costs, Performance and How to Choose in 2026

A few years ago, the choice for most organisations was simple: the best models came from OpenAI, Google and Anthropic through their APIs, and open alternatives were mainly for experimentation. That has changed. Open-weight models from Meta, DeepSeek, Alibaba’s Qwen team, Mistral and, since August 2025, OpenAI itself are capable enough for many production workloads, and they offer control over data, customisation and deployment that closed APIs do not. Closed frontier models still tend to lead on the hardest tasks. This guide explains what “open” actually means, how licences differ, what the evidence says about performance and cost, and how to decide.

What “open” actually means

The Open Source Initiative published the Open Source AI Definition 1.0 in October 2024. To qualify, a system must provide its model weights, the code used to train and run it, and enough information about the training data for a skilled person to substantially recreate it (OSI). Most popular “open” models do not meet that bar, mainly because they do not release full training data. They are more accurately called open-weight models: you can download and run the weights, subject to a licence.

Licences vary considerably:

Model familyTypical licenceKey conditions
Meta LlamaLlama community licenceSeparate licence needed for companies with more than 700 million monthly active users; acceptable use policy (Llama licence)
DeepSeek-R1MITVery permissive, including commercial use and distillation
OpenAI gpt-ossApache 2.0Permissive, without copyleft restrictions (gpt-oss-120b model card)
Qwen and MistralMany models under Apache 2.0Some models carry more restrictive or research-only terms

For businesses, the practical questions are specific: Can we use it commercially? Can we fine-tune it on our data? Can we deploy it in our own cloud or on devices? Can we redistribute a fine-tuned version? The answers differ model by model, so legal review should be part of selection.

How big is the performance gap?

Stanford’s 2025 AI Index reported that open-weight models had narrowed the performance difference with closed models from 8% to just 1.7% on some benchmarks in a single year. That trend has broadly continued, with open models from Chinese labs in particular releasing frequently and competing near the top of public leaderboards.

Two caveats apply. First, benchmark parity does not always translate into parity on hard, multi-step agentic tasks, long-context reliability or safety behaviour, where closed frontier models still tend to lead. Second, the newest closed models often reset the gap when they launch, and open models catch up over the following months. We look at how this plays out in usage and spending in open-weight versus closed models: token volume and spending share.

The cost equation

Per-token prices for open models served by hosting providers are often well below those of frontier closed APIs, and self-hosting can reduce marginal costs further at high, steady volume. Quantisation, which stores weights at lower precision such as 4 bits, lets larger models run on fewer or cheaper GPUs at some cost in quality.

But total cost of ownership includes more than compute:

Cost elementAPISelf-hosted open model
UsagePay per tokenGPU capacity whether used or not
EngineeringMinimalServing, scaling, monitoring, upgrades
Security and complianceVendor’s controls plus your reviewYour responsibility end to end
Model updatesAutomatic, sometimes unwantedYou choose when to upgrade
Utilisation riskNoneIdle capacity at low or spiky volume

Managed providers such as Together AI and Fireworks offer open models through APIs, combining open weights with API convenience. For organisations without an ML operations team, that middle path is often the most economical way to use open models.

When each wins

Open-weight models tend to win when data must stay on your own infrastructure or in a specific jurisdiction; you need to fine-tune deeply on proprietary data; volume is high and steady enough that per-token costs dominate; you need on-device or offline inference; or you want to avoid dependence on a single vendor’s pricing and policies.

Proprietary APIs tend to win when you need the best available performance on complex reasoning, coding or agentic tasks; you lack staff to run infrastructure; usage is low or spiky; you need mature multimodal, voice or tool ecosystems; or you want to prototype quickly without setup.

A framework for choosing

  1. Define the task and quality bar. Build a test set of real examples with known good answers.
  2. Test both kinds of model. Compare a leading closed model with two or three open-weight candidates on accuracy, latency and cost per successful task.
  3. Check the licence and data requirements. Confirm commercial use, fine-tuning and deployment rights, and whether data can leave your environment.
  4. Estimate total cost. Include engineering time and utilisation, not just token prices.
  5. Design for switching. Use an abstraction layer or gateway so you can route tasks to different models and change providers as the market moves.

Many organisations end up with a mix: closed frontier models for high-value, complex tasks and open-weight models for high-volume, sensitive or specialised workloads.

Common misconceptions

“Open means free.” The weights may cost nothing to download, but running them does not. GPUs, engineering time, monitoring and security work all cost money, and some licences restrict commercial use above certain thresholds.

“Self-hosting is automatically more private.” It removes a third party from the data path, but only if your own infrastructure is properly secured. A poorly configured self-hosted endpoint can be riskier than a vendor’s well-audited service.

“Closed APIs lock you in forever.” Lock-in comes mostly from building workflows around one vendor’s proprietary features. Using standard interfaces and keeping prompts, evaluation sets and data portable makes switching far easier, whichever kind of model you start with.

“Open models are unsafe by default.” Safety depends on how a model was trained and how it is deployed. Open models can be fine-tuned to remove safeguards, which is a genuine concern, but organisations deploying them can also add their own filters, monitoring and access controls.

Our analysis

The open-versus-closed decision is no longer about quality alone. Open-weight models are close enough on many benchmarks that licence terms, data control, engineering capacity and total cost often decide the matter. Closed models keep an edge on the hardest tasks and on convenience. The most resilient approach is to test on your own workload, read licences carefully, and build an architecture that can route each job to the model that fits it best.