ctxlen 2026.10.11.dev1__py3-none-any.whl

ctxlen/__init__.py

@@ -0,0 +1,35 @@
+"""max_input and max_output for any litellm model id, from bundled data, with no network."""
+
+import typing as _typing
+
+from ctxlen import _resolve
+
+__all__ = ["max_input", "max_output"]
+
+
+def max_input(model: str) -> _typing.Optional[int]:  # noqa: FA100 (help() shows Optional[int] on 3.9)
+    """The most input tokens `model` accepts, or None if no source knows it.
+
+    `model` is any id you would pass to litellm.completion ("gpt-4o",
+    "anthropic/claude-sonnet-4-5", "bedrock/us.anthropic..."), resolved as
+    litellm resolves it. The number is litellm's; where litellm has none,
+    models.dev's for the same provider's model.
+
+    The sources store what each provider publishes: for some models the input
+    cap (gpt-5: 272,000 of a 400,000 window), for others the whole window,
+    the reply included (Claude). ctxlen does not reconcile them: leave room
+    for the reply.
+    """
+    return _resolve.limits(model, "max_input")[0]
+
+
+def max_output(model: str) -> _typing.Optional[int]:  # noqa: FA100
+    """The most tokens `model` can generate in one reply, or None if no source knows it.
+
+    `model` is any id you would pass to litellm.completion ("gpt-4o",
+    "anthropic/claude-sonnet-4-5", "bedrock/us.anthropic..."), resolved as
+    litellm resolves it. The number is litellm's max_output_tokens for that
+    model; where litellm has none, models.dev's for the same provider's model.
+    An embedding model has no output limit, so it gives None.
+    """
+    return _resolve.limits(model, "max_output")[1]