Open Source vs Closed Source AI Models: What’s the Difference

Every big AI headline this year seems to come from two very different worlds. One sits behind a login screen and a monthly bill. The other lives on GitHub, free for anyone to download and run on their own hardware. I’ve spent weeks testing both types for UK clients, and the gap between them is closing fast. That shift changes what businesses should actually pick in 2026.

What Open Source AI Actually Means

Open source AI models publish their weights — the numbers that make the model work — for anyone to download. Meta’s Llama family, Mistral’s models, and China’s Qwen and DeepSeek all fall into this camp. Once downloaded, you can run them on your own servers with zero ongoing licence fees.

That doesn’t always mean the training data or full code is public too. Many “open weight” models keep their training pipeline private. Over 1.2 million model variants now sit on Hugging Face, most built on a handful of open base models.

UK developers keep asking about this because self-hosting sounds appealing until you price the GPUs. A single Nvidia H100 card still costs upwards of £25,000, and most useful models need several.

What Closed Source AI Actually Means

Closed models — GPT-5.6, Claude, Gemini — never leave the provider’s servers. You send a request through an API and get a response back. The weights, training data, and architecture stay locked away.

This trade-off buys convenience. No GPU procurement, no infrastructure team, no model updates to manage yourself. OpenAI, Anthropic and Google handle scaling, safety tuning and uptime.

The first time I built a client tool on a closed API, the setup took an afternoon. Doing the same with a self-hosted open model took a full week of infrastructure work.

Performance: Has the Gap Actually Closed?

Two years ago closed models dominated every benchmark. That’s no longer true. DeepSeek’s V3 and Qwen’s largest releases now sit within a few points of GPT-5.6 on coding and reasoning tests.

On the LMSYS Chatbot Arena leaderboard, open models occupy roughly a third of the top 20 spots as of mid-2026. That would have been unthinkable in 2024.

Gaps remain. Closed models still lead on tool use, long-context reasoning above 500,000 tokens, and multimodal video understanding. For most everyday tasks, the difference has genuinely narrowed.

Cost: Where the Real Numbers Land

Closed API pricing looks cheap per request. GPT-5.6-mini runs around $0.15 per million input tokens. At low volume that’s nothing.

Scale changes everything. A UK fintech processing 50 million tokens daily can hit £4,000 a month on API fees alone. Self-hosting an equivalent open model on rented cloud GPUs often lands nearer £1,800 monthly.

Break-even usually falls apart before you expect it though. Once you add engineering salaries to maintain the self-hosted stack, small and medium UK businesses rarely come out ahead below a few million requests per month.

Data Privacy and UK Compliance

This is where open models genuinely win for certain sectors. Running a model entirely on your own infrastructure means customer data never leaves the building.

UK healthcare and legal firms I’ve spoken with cite this as the single biggest driver toward self-hosted open models. Under UK GDPR, sending sensitive data to a US-hosted closed API adds a layer of contractual and transfer-risk complexity many firms would rather avoid.

Closed providers have responded with data residency options and no-training guarantees. Microsoft’s Azure OpenAI service now offers UK-region hosting, narrowing this advantage somewhat.

Customisation and Fine-Tuning

Open models let you fine-tune freely. Take Llama, feed it your company’s internal documents, and reshape its behaviour however you like. No permission needed, no usage cap.

Closed providers do offer fine-tuning APIs now, but with limits. OpenAI caps certain fine-tuning features, and the resulting model stays on their servers — you’re renting a customised version, not owning it.

For a UK legal-tech startup I worked with, fine-tuning an open 8-billion parameter model on case law took two days and cost under £200 in compute. That kind of control just isn’t possible with a closed model.

Support, Safety and Reliability

Closed providers carry safety teams, red-teaming units and dedicated abuse-monitoring infrastructure. When something breaks, there’s a support contract and an SLA behind it.

Open models offer none of that by default. Responsibility for content filtering, misuse prevention and uptime shifts entirely onto whoever deploys them. That’s a real burden for smaller teams.

Community support has improved though. Discord servers and GitHub issue trackers for major open models now often resolve bugs faster than enterprise support tickets for closed rivals.

Licensing: The Fine Print Nobody Reads

“Open source” gets used loosely in AI marketing. Llama’s licence restricts commercial use above 700 million monthly active users — fine for almost everyone, but not technically the same as a true open source licence like MIT or Apache.

Mistral’s smaller models ship under genuine Apache 2.0 terms, no strings attached. That distinction matters if a UK startup ever plans to build a product on top of a specific model long-term.

Always check the actual licence text before committing engineering time. A model marketed as “open” can still carry restrictions that closed-source purists would find surprising.

The Environmental and Energy Angle

Running large models, open or closed, burns real electricity. Training a frontier-scale model can consume several gigawatt-hours, roughly equivalent to powering thousands of UK homes for a year.

Closed providers run inference at massive shared scale, which spreads energy cost across millions of users efficiently. Self-hosting a smaller open model locally can actually use less total energy for low-volume UK business use cases.

UK firms weighing sustainability commitments increasingly ask providers for energy-per-query figures. Few closed vendors publish this clearly yet, while some open model maintainers now include rough estimates in their model cards.

Vendor Lock-In: The Hidden Cost of Going Closed

Building an entire product on a single closed API creates dependency few teams think through at the start. If pricing doubles overnight, or the provider changes usage policies, migrating away means rewriting prompts, retesting outputs, and possibly rebuilding whole workflows.

OpenAI raised API pricing tiers twice in 2025 alone. Businesses with no fallback plan absorbed the cost or scrambled to switch providers under time pressure.

Open models sidestep this specific risk. Downloaded weights don’t disappear, don’t change terms overnight, and keep working exactly as they did the day you deployed them, indefinitely.

That stability comes with its own trade-off though. An open model frozen in time can fall behind newer releases, while closed providers push improvements to the same API endpoint automatically.

How UK Regulators Are Treating Both Approaches

The UK AI Safety Institute has published testing frameworks that apply regardless of whether a model is open or closed, focused on capability risks rather than licensing model.

Open weight releases do draw extra scrutiny though, since anyone can fine-tune away safety guardrails once they have the raw weights. Regulators in the US and EU have both flagged this as an emerging concern through 2026.

UK businesses deploying either type of model in regulated sectors — finance, healthcare, legal — still carry full responsibility for outputs regardless of which underlying model powers the product. The regulatory burden doesn’t shift just because a model is free to download.

Real UK Examples Worth Knowing

Stability AI, a UK-founded company, built its reputation on open releases before shifting toward a mixed model with paid enterprise tiers layered on top of open weights. That path is increasingly common among model makers who started fully open.

Several UK banks now run open models internally for document summarisation and fraud pattern analysis, specifically because customer data can’t leave regulated infrastructure. None of that would be possible on a purely closed API under current compliance rules.

Meanwhile, most UK marketing agencies and small consultancies lean almost entirely on closed tools like ChatGPT Enterprise and Claude, since speed to output matters more than infrastructure control for that kind of work.

Ugly workarounds do exist for teams wanting both worlds. Some UK developers now route sensitive queries to a self-hosted open model and everything else to a closed API, splitting traffic based on data sensitivity rather than picking one approach exclusively.

Where This Is Heading in 2027

The line between open and closed keeps blurring rather than hardening. Several closed providers now release smaller “distilled” open versions of their flagship models, partly as marketing, partly as a genuine hedge against the open ecosystem catching up further.

Compute costs keep falling too. Running a capable mid-sized open model that needed a rack of GPUs in 2024 can now run on a single high-end consumer card, putting real self-hosting within reach of far smaller UK teams.

Don’t expect a clean winner. The more realistic outcome is that “open vs closed” stops being the main question, replaced by “which specific model, hosted where, for which specific task” — a far more useful way to think about the decision anyway.

What This Means for You

Pick closed models if you want speed to launch, minimal engineering overhead, and don’t mind ongoing per-token costs. Most small UK businesses testing AI for the first time should start here.

Pick open models if data residency matters, you’re operating at real scale, or you need deep customisation no API will offer. The infrastructure cost is real, but so is the control you get back.

Plenty of UK teams now run both — closed models for customer-facing chat, open models for internal tools handling sensitive data. That hybrid approach is quietly becoming the default.

This article is for educational purposes only and does not constitute financial advice. Cryptocurrency investments involve significant risk. Always do your own research.