The AI supremacy race continues to take Tour de France-style twists. French company Mistral’s latest AI model, Large 4 – colourfully nicknamed Le Chonk – is now the highest scoring ‘open-weight’ model developed outside China, according to one industry benchmark.
Unlike proprietary models such as Anthropic's Claude, OpenAI's GPT and Google's Gemini, an open-weight model gives you the option to download its ‘brain’ – or weights – and run it yourself on your own hardware.
China’s AI models, most of them open-weight, now account for around two-thirds of the AI tokens processed through the market platform OpenRouter.
Particularly given the lower costs of open models, one question we get asked a lot is whether organisations are migrating from closed models to open versions.
The short answer is no. They’re adding, not swapping.
A mix, not a migration
We can’t think of a client that has abandoned closed models entirely. Indeed, recent research from Futurum suggests open-weight models account for just 5% of metered AI consumption.
China’s open models are commonly 60% to 90% cheaper per token than the leading US closed ones, but those headline savings don’t necessarily translate into the total bill. If you choose to self-host an open model, you also take on maintenance and operating costs. Besides, a frontier model might cost more per token but get the job done faster, with fewer attempts and therefore fewer tokens used.
What clients are really buying is optionality. Open and closed models each have their own advantages, and the beauty of a mix is that you don’t have to choose between them. It also gives you some leverage if a provider changes its terms.
AI budgets are still rising, with Gartner expecting $2.7tn of AI spending worldwide this year. Boards are focused less on the price of a token and more on the price of a good outcome – and as importantly, the cost of a bad one.
At PwC, our AI mix is multi-model by design. Our engineering team benchmarks open models continuously, using them in controlled settings.
The right model for the right task
Across our clients, the work being assigned to open models broadly falls into three groups. First, high-volume, well-defined tasks such as extraction, classification, summarisation and search. Secondly, sensitive work, given that businesses can run the model in their own environment and the data doesn’t have to leave it. And, growing fastest, work inside agents and coding tools, where a frontier model plans and checks while open models do the legwork.
Some large banks are going further - adapting open models on their own hardware because this gives them more predictable results and faster response times.
Closed frontier models remain the go-to for the hardest reasoning and agentic work where these models are at the top of their game.
Quality isn’t always the hard part
For straightforward, well-defined tasks, most users wouldn’t notice the difference between an open and a closed model, and most industry trackers put the best open models just months, not years, behind the frontier models.
The hard part is everything around the model. Provenance is the big issue, particularly for regulated clients that must explain where a model came from, which means geopolitics come into play. For this reason, it’s helpful that a number of countries are now developing open-weight models, increasing choice.
There’s also the question of access. It’s easy to take a particular model for granted, but government decisions or changes in supplier terms can leave businesses needing a plan B. In June, Anthropic suspended access to its latest Claude models, Fable 5 and Mythos 5, for foreign nationals after a US export directive.
Having the weights of an open model in your own hands is some insurance against that switch being flipped. Holding the weights isn’t full sovereignty, though. You still need the compute power, and the chips underneath the model can face export controls.
And greater control brings greater responsibility. When you self-host, the model’s built-in safeguards can be removed or changed. That leaves you to put the guardrails in place and keep them working.
There is an upside to that responsibility, beyond control alone. Open models give people the opportunity to get under the bonnet and understand how models behave, how to get the best out of them and where things can go wrong. Without that access, opportunities for this kind of hands-on learning would be far more limited outside the big frontier AI labs. In that sense, open models help democratise AI skills and expertise.
"Nobody I work with is ripping out frontier models. They’re building a mix, with the hardest reasoning on closed models and the high-volume and sensitive work on open ones they control. Trust comes from knowing which model is doing which job, what each outcome costs and who owns the guardrails.”
Jez Bassinder, Chief AI Officer, PwC UKA mix you can trust
Ultimately, the right mix of models will give businesses greater confidence. We recently argued that Britain’s expertise in managing risk and building trust could become a source of AI growth in its own right. The same principle applies here, as understanding the risks and making informed choices about which model to use will help accelerate AI adoption.
But having the right models available is only half the job. You also need to make sure employees are choosing the right one for the task, rather than simply reaching for the newest or most premium option. Going back to the Tour de France analogy, you need to pick the right bike for the challenge – and make sure it has good brakes.
Contact us
Follow us
