diff --git a/tests/entrypoints/openai/completion/test_completion_text_offset.py b/tests/entrypoints/openai/completion/test_completion_text_offset.py
new file mode 100644
index 0000000000..c8ba7df62f
--- /dev/null
+++ b/tests/entrypoints/openai/completion/test_completion_text_offset.py
@@ -0,0 +1,97 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import json
+
+import pytest
+
+from vllm.entrypoints.generate.base.protocol import RequestResponseMetadata
+from vllm.entrypoints.openai.completion.protocol import CompletionRequest
+from vllm.entrypoints.openai.completion.serving import OpenAIServingCompletion
+from vllm.logprobs import Logprob
+from vllm.outputs import CompletionOutput, RequestOutput
+
+TOKENS = [(1, "abc"), (2, " de"), (3, "!")]
+# Text deltas as the detokenizer streams them with stop=["\n\n"]: one char is
+# held back until the request finishes, so the text lags the tokens.
+TEXT_DELTAS = ["ab", "c d", "e!"]
+
+
+def _logprobs(i: int) -> dict[int, Logprob]:
+    token_id, token = TOKENS[i]
+    return {token_id: Logprob(logprob=-1.0, rank=1, decoded_token=token)}
+
+
+def _output(text: str, idxs: list[int], finished: bool) -> RequestOutput:
+    return RequestOutput(
+        request_id="r",
+        prompt="Hi",
+        prompt_token_ids=[9],
+        prompt_logprobs=None,
+        outputs=[
+            CompletionOutput(
+                index=0,
+                text=text,
+                token_ids=[TOKENS[i][0] for i in idxs],
+                cumulative_logprob=None,
+                logprobs=[_logprobs(i) for i in idxs],
+                finish_reason="stop" if finished else None,
+            )
+        ],
+        finished=finished,
+    )
+
+
+def _serving() -> OpenAIServingCompletion:
+    serving = OpenAIServingCompletion.__new__(OpenAIServingCompletion)
+    serving.enable_prompt_tokens_details = False
+    serving.enable_force_include_usage = False
+    serving.enable_per_request_metrics = False
+    serving.system_fingerprint = None
+    serving.return_tokens_as_token_ids = False
+    return serving
+
+
+def _request(stream: bool) -> CompletionRequest:
+    return CompletionRequest(
+        model="m", prompt="Hi", max_tokens=8, logprobs=0, stop=["\n\n"], stream=stream
+    )
+
+
+@pytest.mark.asyncio
+async def test_stream_text_offset_matches_non_stream_with_stop_holdback():
+    async def results():
+        for i, text in enumerate(TEXT_DELTAS):
+            yield 0, _output(text, [i], finished=i == len(TEXT_DELTAS) - 1)
+
+    serving = _serving()
+    stream_offsets = []
+    async for line in serving.completion_stream_generator(
+        _request(stream=True),
+        [{"prompt": "Hi"}],
+        results(),
+        "r",
+        0,
+        "m",
+        num_prompts=1,
+        tokenizer=None,
+        request_metadata=RequestResponseMetadata(request_id="r"),
+    ):
+        data = line.removeprefix("data: ").strip()
+        if data == "[DONE]":
+            continue
+        for choice in json.loads(data)["choices"]:
+            stream_offsets += choice["logprobs"]["text_offset"]
+
+    response = serving.request_output_to_completion_response(
+        [_output("".join(TEXT_DELTAS), [0, 1, 2], finished=True)],
+        _request(stream=False),
+        "r",
+        0,
+        "m",
+        None,
+        RequestResponseMetadata(request_id="r"),
+    )
+    assert response.choices[0].logprobs is not None
+    assert response.choices[0].logprobs.text_offset == [0, 3, 6]
+    assert stream_offsets == [0, 3, 6]
diff --git a/vllm/entrypoints/openai/completion/serving.py b/vllm/entrypoints/openai/completion/serving.py
index 81dd00f0b3..7f7bdac075 100644
--- a/vllm/entrypoints/openai/completion/serving.py
+++ b/vllm/entrypoints/openai/completion/serving.py
@@ -289,7 +289,7 @@ class OpenAIServingCompletion(GenerateBaseServing):
         request_metadata: RequestResponseMetadata,
     ) -> AsyncGenerator[str, None]:
         num_choices = 1 if request.n is None else request.n
-        previous_text_lens = [0] * num_choices * num_prompts
+        next_text_offsets = [0] * num_choices * num_prompts
         previous_num_tokens = [0] * num_choices * num_prompts
         has_echoed = [False] * num_choices * num_prompts
         num_prompt_tokens = [0] * num_prompts
@@ -386,13 +386,18 @@ class OpenAIServingCompletion(GenerateBaseServing):
                             num_output_top_logprobs=request.logprobs,
                             tokenizer=tokenizer,
                             logprob_token_ids=request.logprob_token_ids,
-                            initial_text_offset=previous_text_lens[i],
+                            initial_text_offset=next_text_offsets[i],
                             return_as_token_id=request.return_tokens_as_token_ids,
                         )
+                        # Advance by the tokens just reported, not by delta_text:
+                        # text held back for stop strings lags the tokens.
+                        if logprobs.tokens:
+                            next_text_offsets[i] = logprobs.text_offset[-1] + len(
+                                logprobs.tokens[-1]
+                            )
                     else:
                         logprobs = None
 
-                    previous_text_lens[i] += len(delta_text)
                     previous_num_tokens[i] += len(output.token_ids)
                     stop_reason = output.stop_reason