AI models are becoming outdated at record speed — so what does that mean for the tokens that companies are spending millions on?

AI models are becoming quickly outdated as they’re replaced with newer versions Most AI tokens are used on models available for less than four months Open-weight models are now more popular than closed AI models New data from Vercel’s AI Gateway claims AI models are becoming outdated faster than eve

<![CDATA[ <article> <ul><li><strong>AI models are becoming quickly outdated as they’re replaced with newer versions</strong></li><li><strong>Most AI tokens are used on models available for less than four months</strong></li><li><strong>Open-weight models are now more popular than closed AI models</strong></li></ul><p>New data from Vercel’s AI Gateway claims AI models are becoming outdated faster than ever as new versions are introduced. Token spend is increasingly on newer models – most of which have only been available for the past three months – and as use increases, token price is dropping.</p><p>Simultaneously, Vercel's report suggests the open-weight models are beginning to dominate, with the majority of tokens spent on the models with tunable options over the closed black box systems.</p><p>A key example of this is Fable 5, Anthropic’s most capable model, which lost out to Anthropic’s own Opus 5 as well as OpenAI’s Astra. But upstart models, such as TypeSafe AI’s Jev, are also having an impact.</p><h2 id="token-price-down">Token price down</h2><p>Date from the AI Gateway Production Index shows us that the average price per token has dropped, for the third month in succession, this time by 23.2%. While this is due to various factors, one of the key influences here is the token price way across the industry. Open-weight models make is possible to self-host or run on affordable cloud servers.</p><p>The significance of open-weight models is growing, as enterprise AI use realises that a closed model can be swapped out for a more affordable and capable open-weight alternative. </p><p>Consequently, in the September AI Gateway Production Index, open-weight models have taken 56% of the token volume, but only 14% of the spend on tokens. Open-weight models have grown in popularity throughout 2026, with a token volume share of just 13% back in April and representing just 10% of token use in December 2025.</p><p>Open-weight models making an impact include Meta Llama 4, but Vercel highlights Jev, from TypeSafe AI, as leaving a noticeable footprint in token pricing. The new model has a very low token cost, but relies on a different type of input (data, rather than conversation). They also note it is the “<a href="https://vercel.com/blog/ai-gateway-jev-model-launch" target="_blank">fastest-adopted model</a>” since the AI Gateway Production Index was launched. </p><h2 id="churn-and-spend">Churn and spend</h2><p>While the improvement in AI models means better results for productivity and other uses, the manner of token spend may be about to change. </p><p>With the faster development and release of newer AI models, the commitment to older models – insofar as the purchase of tokens that can only be used on those legacy models – highlights one of the biggest debates in the AI discussion: that AI development is moving too quickly.</p> </article> ]]>