AI models may have a six-week half-life, Vercel data suggests
A Vercel report suggests developers switch AI models roughly every four weeks, with usage fading after about six weeks, according to TechRadar.
According to TechRadar, a new report from Vercel suggests that AI models lose developer attention faster than many people might expect. Vercel's AI Gateway Production Index for November 2026 finds that a typical model reaches its peak use within about four weeks of launch.
TechRadar reports that teams paying for AI services tend to stay with a newly released model for an average of four weeks before moving to another. Vercel's report describes this pattern as a half-life of roughly six weeks, meaning that attention and use decline after that point as developers turn to newer options.
The coverage also points to shifts in which companies are winning new users. According to TechRadar, Anthropic is attracting new teams more effectively than OpenAI, and users show little sign of loyalty to any single provider. Developers are willing to switch models in order to keep their code and AI agents competitive.
The time spent with each new model is shrinking, the outlet notes. The methodology behind the figures is not described in detail, which leaves some questions about how the results were measured.
Background
Modern AI models are large systems trained to generate text, code and other outputs. Companies including OpenAI, Anthropic and Google release new versions regularly, and each release often comes with claims of better reasoning, speed or cost. For developers, the appeal of a fresh model is practical: it may handle a task more accurately or cheaply than the one they already use.
Most teams reach these models through application programming interfaces, or APIs, which let software send requests and receive responses. Some teams also use gateway services that sit between their applications and several providers. Such gateways can make it easier to test or switch between models, because the core application does not have to be rebuilt each time a new option appears.
The phrase half-life comes from physics, where it describes how long it takes for half of a substance to decay. Applied to technology, it suggests that the attention around a product falls over time, usually as something newer appears. Software has long seen waves of interest in frameworks, languages and devices. With AI models, the pace is often measured in weeks rather than years.
Switching costs also matter. Prompts, evaluation tests and workflows are often tuned to a particular model, so moving to another takes effort. Still, if the performance gap between models is large, developers may judge that effort worthwhile. Whether that calculation holds steadily over time is a question that usage data can help answer.
Why it matters
If the pattern described holds, the shelf life of an AI model becomes a planning issue rather than a background detail. Teams that build products around a single model may need to test replacements more often, update documentation and monitor how outputs change. That work takes time, and it can slow features that depend on consistent behavior.
For companies selling models, the finding puts pressure on performance and on the story they tell about it. A model that wins early attention but loses it within weeks has to keep improving to stay relevant. Competitors can gain ground quickly, and newer arrivals can draw users away from established names.
For infrastructure providers, the same trend may increase the value of tools that let teams compare and route between models. If developers expect to switch often, the layer that makes switching easy could become as important as any single model. The report, as described by TechRadar, points toward a market where loyalty must be earned repeatedly rather than assumed.
What to watch
Several developments will show whether the six-week figure is a durable pattern or a snapshot. Future editions of the Vercel index will be an important test, especially if they show similar timing across different periods. Readers should also watch whether Vercel publishes more detail on how its measurements were made, since that would help outside observers judge the claims.
Competitive moves matter too. Model providers may respond with longer-term pricing, stronger developer tools or changes to release schedules. Watch whether gains in new users for any single company hold after the initial launch period, and whether teams begin to standardize on multi-model setups.
Viewpoints
Commentary from our columns: three angles on this story.
The Skeptic open questions and weak points
Every headline number deserves a second look, and this one raises questions before it answers any. A single index built from usage on one platform can tell us what happens on that platform, but it is harder to say how well it reflects the broader market. Teams that route traffic through a gateway may differ from those that call model providers directly, and the sample behind the report is the first thing I would want to understand.
The term half-life also deserves care. It implies a neat decay curve, yet attention to a new model can rise and fall for reasons unrelated to quality: launch hype, pricing promotions, bugs or simple curiosity. A model that fades quickly might be a weak performer, or it might be a fine tool that everyone has already tried. The data alone cannot tell those stories apart.
None of this means the trend is wrong. It means the claim should be tested before anyone rebuilds a strategy around it. Ask what counts as active use, how long the observation window runs, and whether the pattern holds when the market is calm rather than during a crowded release season.
The Ledger the business angle
Loyalty is expensive to earn and cheap to lose, and this report suggests AI companies are learning that lesson in public. If developers stay with a new model for only a few weeks before moving on, pricing power becomes fragile. Each provider must keep shipping improvements, because a lead that lasts a month cannot carry a long-term revenue plan on its own.
The winners may be the companies that make switching painless. Gateways and routing platforms gain value when they let customers test and change models without rewriting code. That position resembles the role of cloud platforms in earlier waves of software, where the infrastructure owner collects steady income regardless of which application wins a given week.
For startups, the message is mixed. A fast-moving market rewards nimble teams that can absorb new models quickly, but it also raises the cost of building anything that depends on one vendor. Investors will likely ask how a business handles developer churn and whether its margins can survive a world in which the favored model changes every few weeks.
The User what changes for people who use it
For most people, the phrase half-life sounds like something for engineers, but the effects reach ordinary users. Many of the AI tools people rely on, from writing assistants to coding helpers, run on models working behind the scenes. When those underlying models change often, the tool you used on Monday may behave differently by Friday, even though the name on the screen stays the same.
That can be useful. Faster turnover means newer capabilities arrive sooner, and people who try many tools may find better fits for specific tasks. It can also be unsettling. Routines built around a particular style of answer may need adjusting, and it becomes harder to know what to trust when the engine underneath keeps shifting.
The practical advice is modest. Keep notes on tasks that matter, check important outputs rather than assuming consistency, and pay attention when a product announces a model change. Treating an AI tool as a changing service, not a fixed product, is a sensible habit for whatever update arrives next.