
Vercel says AI models have a six-week half-life in production

Vercel says in its November 9 Production Index that AI models have a usage half-life of about six weeks on AI Gateway, where more than half of October’s language-model requests went to versions less than 60 days old; the company writes, “Teams are moving to new models almost as fast as labs release them.” The figure describes a shift in requests routed through Vercel’s service, not a loss of capability in older models.
TechRadar’s November 12 account highlights how quickly a typical model reaches peak usage and notes that Anthropic’s Claude Opus 5.5 led estimated gateway spending in its first full month. For an AI startup, the immediate question is whether its product can assess a promising release and redirect traffic without rebuilding the features that depend on its current model.
What the gateway finding measures
AI Gateway sits between applications and model providers, so its traffic shows what participating teams sent through that service. The index aggregates anonymized requests and treats each model version separately. A version’s age begins when it first appears on the gateway, which may differ from the date its developer released it publicly. The findings therefore describe this customer base and this routing path, rather than all production AI use.
LetzView’s November 12 methodology analysis distinguishes the switching cohort: about half of paid teams processing at least 10 million tokens in both September and October changed their most-used model; four in five switchers chose a newer version, and six in ten moved to another lab. Most-used means the model receiving the most tokens. A change in that position shows where the largest share of a team’s work went, not whether its previous choice disappeared from every task.
The gateway also records continued use of older versions, generally at a smaller share of a high-volume team’s traffic. That matters for interpreting the half-life: a team can move its main workload to a recent release while retaining an earlier model for a narrower job. Request counts do not reveal whether any particular move was driven by response quality, latency, price, availability or a change in the application itself.
Why traffic turnover becomes an architecture problem
A model change reaches beyond the name in an API call when product behavior depends on one provider’s conventions. Tool calls, structured responses, context handling and errors can vary across models. If those details are scattered throughout an application, evaluating an alternative requires changes in several product paths. A stable interface around the application’s own tasks can keep those changes contained, while provider adapters handle differences that matter to the result.
That interface should preserve meaningful distinctions. Forcing every model into an identical set of features can hide a capability the product relies on or make failures harder to diagnose. The useful boundary is the product task: what input it accepts, what output is required and which failures are tolerable. A team can then compare candidates against the same requirement, even when the requests sent to each provider differ.
The gateway figures make that design question timely, but they do not decide it for an individual startup. Frequent switching by other customers establishes that alternatives are being tried and adopted. It offers no evidence that the newest version performs a particular product’s task better, or that the work of supporting another provider will repay its maintenance cost.
What makes a model change reversible
A repeatable evaluation supplies the evidence that traffic figures cannot. Representative inputs from the product’s own tasks let a team compare its current model with a candidate on answer quality, failure cases, latency and cost per successful task. Recording the version, configuration and evaluation criteria makes later comparisons intelligible. Without those records, an apparent improvement may reflect a different prompt, workload or scoring rule.
Production routing needs a way back as well as a way forward. A limited rollout can expose changes in output shape, tool behavior or error rates before all requests move. Keeping the prior route available allows the team to restore a working version if the candidate causes a problem under live conditions. This is a product control: users experience the response, regardless of which provider generated it.
Cost belongs in the same comparison. The index estimates spending from published list prices, while a startup’s actual cost depends on its mix of inputs, outputs, retries and contracted rates. A cheaper token price may deliver little savings if a task needs more tokens or more attempts to succeed. Monitoring cost alongside task outcomes makes a change in routing a measurable operating decision.
When model portability earns its upkeep
Portability is most valuable when a product has important workloads for which alternative models repeatedly produce a material gain, or when dependence on a single provider makes a disruption costly. In those cases, maintaining adapters, evaluations and reversible routing can be justified by the choices they make possible. A stable workload with no demonstrated benefit from switching may need only a clear boundary around its current provider, leaving room to expand that boundary when evidence warrants it.
The consequence of faster turnover is a shorter interval between opportunities to reconsider a model choice, not a deadline to replace it. For a startup, the durable asset is the ability to compare a new release with the version already serving users, then move traffic only when the measured product result warrants the change.
Also read:
Related articles


QUASA is heading to Blockchain Life 2026 — let’s connect in Dubai on December 1–2!

Pinecone vs Qdrant: Usage Billing or Reserved Compute Decides the Cost

Nearly Half of Required AI Users Resist It—Wellness May Explain Why

Mercury 3 Claims 1,068 Tokens per Second—Independent Tests Find Trade-offs

Choose RAG Chunk Size by Answer Length—not a Universal Token Count
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.