One model string moves Vercel AI Gateway traffic to Claude Sonnet 5.5

According to the review, one model string handles version management across the client libraries a team already runs ^[4]. It says that job used to take four or five separate client libraries and credential stores ^[14]. An identical API surface keeps Sonnet 5 requests valid ^[11]. Whether the new model's answers still pass the checks downstream of those requests is a separate question, and the review's own migration step allows for it: "swap the model ID, run your test suite, ship it" ^[12].

The Terminal-Bench gap works out to 60.3 points, or about 6.9 times Sonnet 5's score ^[9]. The review describes the benchmark as a measure of multi-step agentic task completion in terminal environments ^[1]. For that number to carry over, a workload has to look like the benchmark: long chains of tool calls where one early mistake sinks the task. The review goes further. It recommends moving coding, document generation and routine agentic work off Sonnet 5 and spending the freed budget on Opus 5.5 ^[7]. Document generation is not a terminal-agent task, so a terminal score says little about it.

The cost claim needs its own check. Token pricing is identical to Sonnet 5 ^[3], yet the review reports roughly 30% more throughput per dollar in practice ^[10]. At a fixed price per token, a faster model finishes sooner and bills the same. Getting 30% more work per dollar means each task uses about 77% of its old token count, roughly 23% fewer tokens ^[8]. The review points to where the saving could come from: multi-step tasks that no longer have to be split into sequential calls to keep context manageable ^[13].

The review aims Zero Data Retention at healthcare, finance and "any environment where data residency isn't optional" ^[2]. Retention covers whether a provider keeps request data. Residency covers where that data is processed. A zero-retention setting meets only the first of those requirements ^[2]. The review also does not say whether ZDR is on by default, set per request or set per account, or whether it applies on all three endpoints ^[5]. In my view, a regulated team should get that setting in writing from Vercel before it points production traffic at the new ID.

The model swap changes only the model ID ^[11]. The credential change that week is a separate product. With Vercel Connect, a team registers a connector once per provider and calls getToken() at request time. Connect then uses the deployment's OIDC identity to issue an ephemeral, scoped credential ^[15]. No bot token sits in an environment variable, and no shared secret has to be rotated across environments ^[15]. Tying credentials to the deployment's identity is sound engineering for agents that reach into Slack, GitHub, databases and internal APIs. The review says credential management for those agents becomes a rotation and audit liability at scale ^[18].