ctxlen 2026.10.11.dev3__tar.gz → 2026.10.12.dev1__tar.gz

ctxlen-2026.10.12.dev1/PKG-INFO

@@ -0,0 +1,65 @@
+Metadata-Version: 2.4
+Name: ctxlen
+Version: 2026.10.12.dev1
+Summary: The context limits of any LLM: always fresh, works offline, three functions
+Keywords: llm,litellm,context window,tokens,models.dev
+License-Expression: MIT
+License-File: LICENSE
+Classifier: Programming Language :: Python :: 3
+Classifier: Typing :: Typed
+Requires-Dist: httpx>=0.28,<1
+Requires-Dist: certifi
+Requires-Python: >=3.9
+Project-URL: Repository, https://github.com/michaeltheologitis/ctxlen
+Description-Content-Type: text/markdown
+
+# ctxlen 📏
+
+A super minimal, always-fresh, project to look up the context limits (input and output tokens) of LLMs. 🤔
+
+- 🔄 New models checked every 6 hours, and your install picks them up within the hour: no upgrade needed.
+- 🪶 Just httpx and certifi as dependencies, and it still works offline, with the numbers it shipped with.
+- 🧰 Just three functions.
+
+## Install
+
+```shell
+uv add ctxlen
+```
+
+Or with pip:
+
+```shell
+pip install ctxlen
+```
+
+## Usage
+
+Ask for a model's whole context window:
+
+```python
+import ctxlen
+
+ctxlen.ctxlen("claude-opus-5-5")  # 1_000_000
+ctxlen.ctxlen("gpt-6-sol")        # 1_050_000
+```
+
+You can also put the provider in front, like litellm does:
+
+```python
+ctxlen.ctxlen("anthropic/claude-sonnet-5-5")  # 1_000_000
+ctxlen.ctxlen("openai/gpt-6-luna")            # 1_050_000
+```
+
+Or ask for the input and output limits on their own:
+
+```python
+ctxlen.max_input("gpt-6-sol")   # 922_000
+ctxlen.max_output("gpt-6-sol")  # 128_000
+```
+
+The numbers come from litellm and models.dev; their notices are in [LICENSE](https://github.com/michaeltheologitis/ctxlen/blob/main/LICENSE).
+
+## Offline
+
+`ctxlen.ctxlen("gpt-6-sol", live=False)` uses only the numbers that came with your version, and `CTXLEN_OFFLINE=1` does that for the whole process. ctxlen asks GitHub for fresh numbers at most once an hour and keeps them in `~/.cache/ctxlen/`.

ctxlen-2026.10.12.dev1/README.md

@@ -0,0 +1,50 @@
+# ctxlen 📏
+
+A super minimal, always-fresh, project to look up the context limits (input and output tokens) of LLMs. 🤔
+
+- 🔄 New models checked every 6 hours, and your install picks them up within the hour: no upgrade needed.
+- 🪶 Just httpx and certifi as dependencies, and it still works offline, with the numbers it shipped with.
+- 🧰 Just three functions.
+
+## Install
+
+```shell
+uv add ctxlen
+```
+
+Or with pip:
+
+```shell
+pip install ctxlen
+```
+
+## Usage
+
+Ask for a model's whole context window:
+
+```python
+import ctxlen
+
+ctxlen.ctxlen("claude-opus-5-5")  # 1_000_000
+ctxlen.ctxlen("gpt-6-sol")        # 1_050_000
+```
+
+You can also put the provider in front, like litellm does:
+
+```python
+ctxlen.ctxlen("anthropic/claude-sonnet-5-5")  # 1_000_000
+ctxlen.ctxlen("openai/gpt-6-luna")            # 1_050_000
+```
+
+Or ask for the input and output limits on their own:
+
+```python
+ctxlen.max_input("gpt-6-sol")   # 922_000
+ctxlen.max_output("gpt-6-sol")  # 128_000
+```
+
+The numbers come from litellm and models.dev; their notices are in [LICENSE](https://github.com/michaeltheologitis/ctxlen/blob/main/LICENSE).
+
+## Offline
+
+`ctxlen.ctxlen("gpt-6-sol", live=False)` uses only the numbers that came with your version, and `CTXLEN_OFFLINE=1` does that for the whole process. ctxlen asks GitHub for fresh numbers at most once an hour and keeps them in `~/.cache/ctxlen/`.

{ctxlen-2026.10.11.dev3 → ctxlen-2026.10.12.dev1}/pyproject.toml

@@ -1,9 +1,12 @@
 [project]
 name = "ctxlen"
-version = "2026.10.11.dev3"
-description = "max_input and max_output for any litellm model id, with no dependencies and no network"
+version = "2026.10.12.dev1"
+description = "The context limits of any LLM: always fresh, works offline, three functions"
 requires-python = ">=3.9"
-dependencies = []
+dependencies = [
+    "httpx>=0.28,<1",
+    "certifi",
+]
 license = "MIT"
 license-files = ["LICENSE"]
 readme = "README.md"

{ctxlen-2026.10.11.dev3 → ctxlen-2026.10.12.dev1}/pyproject.toml.orig

@@ -1,9 +1,9 @@
 [project]
 name = "ctxlen"
-version = "2026.10.11.dev3"
-description = "max_input and max_output for any litellm model id, with no dependencies and no network"
+version = "2026.10.12.dev1"
+description = "The context limits of any LLM: always fresh, works offline, three functions"
 requires-python = ">=3.9"
-dependencies = []
+dependencies = ["httpx>=0.28,<1", "certifi"]
 license = "MIT"
 license-files = ["LICENSE"]
 readme = "README.md"

ctxlen-2026.10.12.dev1/src/ctxlen/__init__.py

@@ -0,0 +1,64 @@
+"""The context limits of any litellm model id: the latest release's numbers, fetched from GitHub
+at most once an hour, else the ones installed with this version."""
+
+import typing as _typing
+
+from ctxlen import _resolve
+
+__all__ = ["ctxlen", "max_input", "max_output"]
+
+
+def ctxlen(model: str, *, live: bool = True) -> _typing.Optional[int]:  # noqa: FA100
+    """The total context window, input and output tokens together, or None.
+
+    `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 models.dev's for the same provider's
+    model; where models.dev has none, or one below this model's limits, it is
+    max_input(model), raised to max_output(model) where that is larger,
+    which is never larger than the true total.
+
+    With live=True, the default, the numbers are the latest release's, fetched
+    from GitHub at most once an hour and cached in ~/.cache/ctxlen/. Offline,
+    with live=False, or with CTXLEN_OFFLINE=1 set, they are the ones installed
+    with this version.
+    """
+    return _resolve.limits(model, "ctxlen", live)[2]
+
+
+def max_input(model: str, *, live: bool = True) -> _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.
+
+    With live=True, the default, the numbers are the latest release's, fetched
+    from GitHub at most once an hour and cached in ~/.cache/ctxlen/. Offline,
+    with live=False, or with CTXLEN_OFFLINE=1 set, they are the ones installed
+    with this version.
+    """
+    return _resolve.limits(model, "max_input", live)[0]
+
+
+def max_output(model: str, *, live: bool = True) -> _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.
+
+    With live=True, the default, the numbers are the latest release's, fetched
+    from GitHub at most once an hour and cached in ~/.cache/ctxlen/. Offline,
+    with live=False, or with CTXLEN_OFFLINE=1 set, they are the ones installed
+    with this version.
+    """
+    return _resolve.limits(model, "max_output", live)[1]