Assistant-turn prefill is not supported on this model

The pattern that stopped working

Prefilling the assistant turn — starting the model's response with text you've already written, so it continues from there rather than starting cold — is a long-standing technique for forcing a specific output format or skipping past a predictable preamble. On a model where thinking is on by default or cannot be turned off, that same request now gets rejected.

Why prefill and thinking don't mix

The mechanism a model uses to think before it answers has to run before any text — including whatever you prefilled — is produced, and a prefilled assistant turn assumes the opposite: that generation starts exactly where your text leaves off. Those two assumptions are structurally incompatible on a model that reasons first, not just awkward to combine, which is why the newer generation of models with default-on or always-on thinking rejects a prefilled assistant turn rather than silently dropping the thinking step to accommodate it. Silently dropping thinking to honour a prefill would be a worse outcome than an explicit error, since it would change the model's behaviour without telling you.

Getting the outcome you actually wanted without prefill

Most uses of prefill were really trying to enforce output shape — always starting with a specific tag, skipping a conversational preamble, forcing valid JSON. On a thinking-capable model, get the same result through the system prompt and, where available, structured-output constraints instead of the assistant turn itself. It's a different mechanism, but it targets the same goal — shape, not content — and doesn't collide with how these models reason.

If you need prefill specifically

If prefill genuinely is the right tool for your use case rather than a workaround for something structured outputs would solve better, that only works on a model where thinking is off by default and can stay off. Whether that's viable depends on the same trade the rest of that model's thinking behaviour involves — see what actually changed from Opus 4.6 to Opus 5 for which models default thinking off versus on, since that default is what actually decides whether this error hits you.