Time to rethink US frontier AI model dominance

China has outdone the US again, with benchmarks from Artificial Analysis showing that the latest DeepSeek model, V4 Flash 0731, is significantly cheaper than equivalent OpenAI models. Even after OpenAI’s most recent 80% price cut on GPT-5.6 Luna, Artificial Analysis reported that DeepSeek V4 Flash 0731’s cost per task using its first-party application programming interface (API) is 60% lower than GPT-5.6 According to Artificial Analysis, DeepSeek’s price performance lead is due to the significant cost savings that can be achieved when the answer to a prompt is already cached. Artificial Analysis reported that DeepSeek offers a cache hit discount rate of about 98% on its first-party API, which it said is a “significantly more aggressive discount than the 90% cache hit discount offered by most of the industry”. Beyond the significant cost savings available from the latest DeepSeek option, IT decision-makers assessing which model to choose may find that there are benefits to avoiding the frontier models, which come at a premium. This is reflected in the findings of the recently published AI openness report from OpenUK. Discussing the report, Amanda Brock, CEO of OpenUK, said: “The data tells us that the UK is not well served by throwing money at building new large or frontier models. The opportunity to home in on with a laser focus is the model-adjacent technologies that AI harnesses – infrastructure and developer tooling . The lower barrier to entry is enabling individuals and innovators to break into this space.” Another consideration is the rate of change, which means that some newer models outperform frontier AI models that were released just a few months ago, as Sapien CTO Glen Ceniza explains following a meeting he had with Anthropic. “One thing they told me is that their ‘midweight’ model now is significantly better than the heavyweight model from three months ago,” he said. In other words, Asana, which uses Anthropic’s Claude model, today outperforms Opus from three months ago. “But people have it in their mind that they want to use Opus even if the midweight model today is significantly better, faster and uses less tokens. You’ll get better results than the older heavyweight model.” Opting for the less expensive non-frontier AI models or using open models gives IT leaders greater choice when they need to decide how to budget for their organisation’s AI strategy and balance the needs of the organisation with IT and AI costs. For instance, housing association Sovereign Network Group (SNG) has decided to be AI model agnostic. Vanessa Hewitt, its head of business platforms, said: “We have to be agnostic because the price is constantly changing.” By having what she says is a “headless model”, SNG focuses its technology investments on the foundational technology for integrating IT systems and providing access to the data across the organisation. Boomi’s integration platform is one of the pieces of this foundational technology. SNG’s goal is to provide affordable housing in the South East and London, and she says Boomi supports this objective: As an example, she says: “Boomi helps us in initiatives where we can offer a customer a third-party app to help them with rent arrears.” For Hewitt, “AI is peripheral”. She says it supports the niche use cases SNG wants to do. Read more about AI cost management Boomi CEO shares vision of AI cost management : Steve Lucas, CEO of Boomi, believes the answer is prompt routing, which sends queries to the LLM with the lowest token cost and caches responses. AI coding agents will cost more than real developers: As organisations ramp up the use of AI coding agents in software development, they may find costs increase significantly if such tools are overused.