Claude Opus 5.5 Arrives With API Changes That Break Some Requests

Anthropic's Claude Opus 5.5 is now available through Vercel's AI Gateway, a service that lets developers call models from several providers through one API, according to Vercel's changelog. Vercel describes the model as a step-change improvement over Opus 5, with its biggest gains in agentic coding, long-running agent tasks and knowledge work.

The headline performance claim comes from Anthropic, relayed by Vercel: Opus 5.5 performs at the level of Claude Fable 5.1 while running about 30% faster and costing about 40% less per task than Opus 5. Those are the vendor's own figures, and Vercel's post cites no independent benchmark behind them.

The more immediate consequence for developers is two API changes that Vercel warns can turn previously valid requests into HTTP 400 errors. First, thinking is always adaptive: the model decides how much to reason on each request, and requests that disable thinking or set a fixed thinking budget are rejected. Vercel's guidance is to steer reasoning with the effort setting and with prompting instead.

Second, forced tool use is retired. A request can no longer require the model to call a tool, or to call a specific one. Vercel recommends prompting the model toward the tool and having the calling harness catch and retry misses, and says teams that forced a tool call to get JSON back should switch to structured outputs.

Together these changes mean the upgrade is not a drop-in model swap for every integration. Code that forces a tool call or pins a thinking budget has to be changed before it moves to Opus 5.5, while code that already relies on adaptive thinking and ordinary tool prompting should be unaffected by these two rules.

Per Vercel, the model has a 1 million token context window, returns up to 128,000 output tokens and has a June 2026 knowledge cutoff. It is addressable as anthropic/claude-opus-5.5 through the AI SDK, OpenAI-compatible Chat Completions, the Anthropic Messages API and connected coding agents, with a fast mode available through a gateway speed option. Vercel also says the model reports back in plain language on what it did, what it found and what it needs next, which it presents as making long runs easier to supervise.