“This model does not support assistant message prefill” — which Claude models dropped prefill, and what replaces it

Which Claude models reject prefill (as of September 2026)

Anthropic's error reference is blunt: Claude 4.6 and later models do not support prefilling the last assistant message, and a request that does gets exactly this 400 (https://platform.claude.com/docs/en/api/errors#prefill-not-supported). By name, that is Claude Opus 4.6 and every later Opus, Claude Opus 5.5 included (https://platform.claude.com/docs/en/models/opus-5-5/migration-guide); Claude Sonnet 4.6 and Claude Sonnet 5 (https://platform.claude.com/docs/en/models/sonnet-5/migration-guide); and Claude Fable 5 and Fable 5.1 (https://platform.claude.com/docs/en/models/fable-5-1/migration-guide). Claude Haiku 4.5 still accepts a prefill, as do Claude Sonnet 4.5 and Claude Opus 4.5 — which is why this error tends to arrive with a model-ID change rather than a code change. Assistant messages earlier in the conversation, few-shot examples included, are not affected.

End on a user turn — including the one-line fix

Often nobody wrote a prefill on purpose: something in the stack put an assistant message last. Public reports trace it to a memory manager that appends one (Strands issue #1694), an agent loop that re-requests after a turn it wrongly thinks is unfinished (opencode issue #46415), back-to-back replies and agent handoffs (LiveKit issue #4907), and an empty assistant message left at the end (AutoGen PR #7931). If the trailing message is empty, drop it — that is the one line. If it holds text you want continued, do what Anthropic's migration guidance says for continuations: put the continuation in a user message that quotes where the answer stopped. One documented exception to leave alone: a pause_turn from server tools such as web search is continued by sending the paused assistant content back as-is (https://platform.claude.com/docs/en/agents-and-tools/tool-use/server-tools).

from openai import OpenAI

client = OpenAI(base_url="https://api.kunavo.com/v1", api_key="sk-kn-...")

def end_on_user(messages: list[dict]) -> list[dict]:
    """Claude 4.6 and later reject a conversation whose last turn is the assistant's."""
    last = messages[-1] if messages else None
    if not last or last["role"] != "assistant" or last.get("tool_calls"):
        return messages               # tool_calls are owed tool results instead
    content = last.get("content")
    if not content or (isinstance(content, str) and not content.strip()):
        return messages[:-1]          # the one line: drop an empty tail
    tail = content[-200:] if isinstance(content, str) else "..."
    return messages + [{
        "role": "user",
        "content": f"Your previous response was interrupted and ended with {tail!r}. "
                   "Continue from where you left off.",
    }]

messages = [
    {"role": "user", "content": "Explain HTTP caching in three bullets."},
    {"role": "assistant", "content": ""},  # the empty tail a framework left behind
]

resp = client.chat.completions.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=end_on_user(messages),
)
print(resp.choices[0].message.content)

Replace what the prefill was doing

Anthropic's prompting guide maps each old use of prefill to a replacement (https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices#migrating-away-from-prefilled-responses). Forcing JSON: structured outputs, which constrain the response to your schema and are available on Claude Haiku 4.5, Sonnet 4.5, Opus 4.5 and every model since (https://platform.claude.com/docs/en/build-with-claude/structured-outputs); for YAML or another format, the guide's advice is to ask for the structure and retry when it misses. Classification: a tool with an enum of your valid labels, or structured outputs. Skipping a preamble: a system-prompt instruction such as “Respond directly without preamble.” Holding a persona: set the role in the system prompt (https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices#give-claude-a-role), and put periodic reminders in the user turn instead of a prefilled assistant one. A cut-off answer: the user-turn continuation above. Forcing a tool call is not a drop-in substitute everywhere — Claude Fable 5.1 and Claude Opus 5.5 reject tool_choice any and tool with a 400 of their own (https://platform.claude.com/docs/en/api/errors#forced-tool-use-not-supported).

# Before: messages ended with {"role": "assistant", "content": "{"}
# Through Kunavo, parse and check the JSON that comes back (see below).
curl https://api.kunavo.com/v1/messages \
  -H "x-api-key: sk-kn-..." \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-4-6",
    "max_tokens": 1024,
    "messages": [
      {"role": "user", "content": "Extract the name and email: John Smith <john@example.com>"}
    ],
    "output_config": {
      "format": {
        "type": "json_schema",
        "schema": {
          "type": "object",
          "properties": {
            "name": {"type": "string"},
            "email": {"type": "string"}
          },
          "required": ["name", "email"],
          "additionalProperties": false
        }
      }
    }
  }'

Which Claude models don't support assistant message prefill?

Per Anthropic's docs as of September 2026, every Claude model from 4.6 on: Claude Opus 4.6 and every later Opus, Claude Sonnet 4.6 and Sonnet 5, Claude Fable 5 and 5.1, and the Mythos models. Claude Haiku 4.5, Sonnet 4.5 and Opus 4.5 still accept one.

How do I force JSON output from Claude without prefill?

Use structured outputs: pass a JSON schema in output_config.format and the response is constrained to it. Anthropic lists message prefilling as incompatible with JSON outputs in any case.

Why do I get this error when I never prefilled anything?

Something in your stack left an assistant message last — a memory or session manager, an agent loop re-requesting after a turn, an agent handoff, or an empty assistant message nobody removed. Log the messages array exactly as it is sent; its last element will have role: "assistant".

Can I still include assistant messages in the conversation?

Yes. Only the final turn is restricted; earlier assistant messages, few-shot examples included, are unaffected.

Why does it happen through an OpenAI-compatible endpoint too?

Because the gateway carries your final role: "assistant" message across as a final Claude assistant turn — the same prefill, reshaped. Kunavo's chat translator does exactly that, and the same message has been reported through other OpenAI-compatible gateways (opencode issue #13768). The fix is the same: end on a user message.