Skip to content

Framework for orchestrating role-playing, autonomous AI agents. By fostering collaborative intelligence, CrewAI empowers agents to work together seamlessly, tackling complex tasks.

refactor(llms): centralize and refresh context windows (#7796)

Vidit Ostwala6e6d0f9d85a
diff --git a/lib/crewai/src/crewai/llm.py b/lib/crewai/src/crewai/llm.py
index c352417b4..569296bb8 100644
--- a/lib/crewai/src/crewai/llm.py
+++ b/lib/crewai/src/crewai/llm.py
@@ -1,7 +1,7 @@
from __future__ import annotations
from collections import defaultdict
-from collections.abc import Callable
+from collections.abc import Callable, Mapping
from datetime import datetime
import json
import logging
@@ -48,6 +48,13 @@ from crewai.llms.constants import (
GEMINI_MODELS,
OPENAI_MODELS,
)
+from crewai.llms.context_window import (
+ AZURE_OPENAI_CONTEXT_WINDOWS,
+ DEFAULT_CONTEXT_WINDOW_SIZE,
+ LLM_CONTEXT_WINDOW_SIZES,
+ OPENAI_CONTEXT_WINDOWS,
+ resolve_context_window_size,
+)
from crewai.utilities import InternalInstructor
from crewai.utilities.exceptions.context_window_exceeding_exception import (
LLMContextLengthExceededError,
@@ -164,171 +171,8 @@ def _ensure_litellm() -> bool:
return LITELLM_AVAILABLE
-MIN_CONTEXT: Final[int] = 1024
-MAX_CONTEXT: Final[int] = 2097152 # Current max from gemini-1.5-pro
ANTHROPIC_PREFIXES: Final[tuple[str, str, str]] = ("anthropic/", "claude-", "claude/")
-LLM_CONTEXT_WINDOW_SIZES: Final[dict[str, int]] = {
- "gpt-4": 8192,
- "gpt-4o": 128000,
- "gpt-4o-mini": 128000,
- "gpt-5.4-mini": 200000,
- "gpt-5.6": 1050000, # sol, terra, luna, and the gpt-5.6 alias
- "gpt-4-turbo": 128000,
- "gpt-4.1": 1047576, # Based on official docs
- "gpt-4.1-mini-2025-04-14": 1047576,
- "gpt-4.1-nano-2025-04-14": 1047576,
- "o1-preview": 128000,
- "o1-mini": 128000,
- "o3-mini": 200000,
- "o4-mini": 200000,
- "gemini-3.8-flash": 1048576,
- "gemini-3-pro-preview": 1048576,
- "gemini-2.0-flash": 1048576,
- "gemini-2.0-flash-thinking-exp-01-21": 32768,
- "gemini-2.0-flash-lite-001": 1048576,
- "gemini-2.0-flash-001": 1048576,
- "gemini-2.5-flash-preview-04-17": 1048576,
- "gemini-2.5-pro-exp-03-25": 1048576,
- "gemini-1.5-pro": 2097152,
- "gemini-1.5-flash": 1048576,
- "gemini-1.5-flash-8b": 1048576,
- "gemini/gemma-3-1b-it": 32000,
- "gemini/gemma-3-4b-it": 128000,
- "gemini/gemma-3-12b-it": 128000,
- "gemini/gemma-3-27b-it": 128000,
- "deepseek-chat": 128000,
- "gemma2-9b-it": 8192,
- "gemma-7b-it": 8192,
- "llama3-groq-70b-8192-tool-use-preview": 8192,
- "llama3-groq-8b-8192-tool-use-preview": 8192,
- "llama-3.1-70b-versatile": 131072,
- "llama-3.1-8b-instant": 131072,
- "llama-3.2-1b-preview": 8192,
- "llama-3.2-3b-preview": 8192,
- "llama-3.2-11b-text-preview": 8192,
- "llama-3.2-90b-text-preview": 8192,
- "llama3-70b-8192": 8192,
- "llama3-8b-8192": 8192,
- "mixtral-8x7b-32768": 32768,
- "llama-3.3-70b-versatile": 128000,
- "llama-3.3-70b-instruct": 128000,
- "Meta-Llama-3.3-70B-Instruct": 131072,
- "QwQ-32B-Preview": 8192,
- "Qwen2.5-72B-Instruct": 8192,
- "Qwen2.5-Coder-32B-Instruct": 8192,
- "Meta-Llama-3.1-405B-Instruct": 8192,
- "Meta-Llama-3.1-70B-Instruct": 131072,
- "Meta-Llama-3.1-8B-Instruct": 131072,
- "Llama-3.2-90B-Vision-Instruct": 16384,
- "Llama-3.2-11B-Vision-Instruct": 16384,
- "Meta-Llama-3.2-3B-Instruct": 4096,
- "Meta-Llama-3.2-1B-Instruct": 16384,
- "us.amazon.nova-pro-v1:0": 300000,
- "us.amazon.nova-micro-v1:0": 128000,
- "us.amazon.nova-lite-v1:0": 300000,
- "us.anthropic.claude-opus-4-7": 1000000,
- "us.anthropic.claude-sonnet-4-6": 1000000,
- "us.anthropic.claude-opus-4-6-v1": 1000000,
- "us.anthropic.claude-opus-4-5-20251101-v1:0": 200000,
- "us.anthropic.claude-haiku-4-5-20251001-v1:0": 200000,
- "us.anthropic.claude-sonnet-4-5-20250929-v1:0": 200000,
- "us.anthropic.claude-opus-4-1-20250805-v1:0": 200000,
- "us.anthropic.claude-opus-4-20250514-v1:0": 200000,
- "us.anthropic.claude-sonnet-4-20250514-v1:0": 200000,
- "us.anthropic.claude-3-5-sonnet-20240620-v1:0": 200000,
- "us.anthropic.claude-3-5-haiku-20241022-v1:0": 200000,
- "us.anthropic.claude-3-5-sonnet-20241022-v2:0": 200000,
- "us.anthropic.claude-3-7-sonnet-20250219-v1:0": 200000,
- "us.anthropic.claude-3-sonnet-20240229-v1:0": 200000,
- "us.anthropic.claude-3-opus-20240229-v1:0": 200000,
- "us.anthropic.claude-3-haiku-20240307-v1:0": 200000,
- "us.meta.llama3-2-11b-instruct-v1:0": 128000,
- "us.meta.llama3-2-3b-instruct-v1:0": 131000,
- "us.meta.llama3-2-90b-instruct-v1:0": 128000,
- "us.meta.llama3-2-1b-instruct-v1:0": 131000,
- "us.meta.llama3-1-8b-instruct-v1:0": 128000,
- "us.meta.llama3-1-70b-instruct-v1:0": 128000,
- "us.meta.llama3-3-70b-instruct-v1:0": 128000,
- "us.meta.llama3-1-405b-instruct-v1:0": 128000,
- "eu.anthropic.claude-3-5-sonnet-20240620-v1:0": 200000,
- "eu.anthropic.claude-3-sonnet-20240229-v1:0": 200000,
- "eu.anthropic.claude-3-haiku-20240307-v1:0": 200000,
- "eu.anthropic.claude-opus-4-7": 1000000,
- "eu.anthropic.claude-sonnet-4-6": 1000000,
- "eu.anthropic.claude-opus-4-6-v1": 1000000,
- "eu.anthropic.claude-opus-4-5-20251101-v1:0": 200000,
- "eu.anthropic.claude-haiku-4-5-20251001-v1:0": 200000,
- "eu.anthropic.claude-sonnet-4-5-20250929-v1:0": 200000,
- "eu.anthropic.claude-opus-4-1-20250805-v1:0": 200000,
- "eu.anthropic.claude-opus-4-20250514-v1:0": 200000,
- "eu.anthropic.claude-sonnet-4-20250514-v1:0": 200000,
- "eu.meta.llama3-2-3b-instruct-v1:0": 131000,
- "eu.meta.llama3-2-1b-instruct-v1:0": 131000,
- "apac.anthropic.claude-3-5-sonnet-20240620-v1:0": 200000,
- "apac.anthropic.claude-3-5-sonnet-20241022-v2:0": 200000,
- "apac.anthropic.claude-3-sonnet-20240229-v1:0": 200000,
- "apac.anthropic.claude-3-haiku-20240307-v1:0": 200000,
- "apac.anthropic.claude-opus-4-7": 1000000,
- "apac.anthropic.claude-sonnet-4-6": 1000000,
- "apac.anthropic.claude-opus-4-6-v1": 1000000,
- "apac.anthropic.claude-opus-4-5-20251101-v1:0": 200000,
- "apac.anthropic.claude-haiku-4-5-20251001-v1:0": 200000,
- "apac.anthropic.claude-sonnet-4-5-20250929-v1:0": 200000,
- "apac.anthropic.claude-opus-4-1-20250805-v1:0": 200000,
- "apac.anthropic.claude-opus-4-20250514-v1:0": 200000,
- "apac.anthropic.claude-sonnet-4-20250514-v1:0": 200000,
- "amazon.nova-pro-v1:0": 300000,
- "amazon.nova-micro-v1:0": 128000,
- "amazon.nova-lite-v1:0": 300000,
- "anthropic.claude-opus-4-7": 1000000,
- "anthropic.claude-sonnet-4-6": 1000000,
- "anthropic.claude-opus-4-6-v1": 1000000,
- "anthropic.claude-opus-4-5-20251101-v1:0": 200000,
- "anthropic.claude-haiku-4-5-20251001-v1:0": 200000,
- "anthropic.claude-sonnet-4-5-20250929-v1:0": 200000,
- "anthropic.claude-opus-4-1-20250805-v1:0": 200000,
- "anthropic.claude-opus-4-20250514-v1:0": 200000,
- "anthropic.claude-sonnet-4-20250514-v1:0": 200000,
- "anthropic.claude-3-5-sonnet-20240620-v1:0": 200000,
- "anthropic.claude-3-5-haiku-20241022-v1:0": 200000,
- "anthropic.claude-3-5-sonnet-20241022-v2:0": 200000,
- "anthropic.claude-3-7-sonnet-20250219-v1:0": 200000,
- "anthropic.claude-3-sonnet-20240229-v1:0": 200000,
- "anthropic.claude-3-opus-20240229-v1:0": 200000,
- "anthropic.claude-3-haiku-20240307-v1:0": 200000,
- "anthropic.claude-v2:1": 200000,
- "anthropic.claude-v2": 100000,
- "anthropic.claude-instant-v1": 100000,
- "meta.llama3-1-405b-instruct-v1:0": 128000,
- "meta.llama3-1-70b-instruct-v1:0": 128000,
- "meta.llama3-1-8b-instruct-v1:0": 128000,
- "meta.llama3-70b-instruct-v1:0": 8000,
- "meta.llama3-8b-instruct-v1:0": 8000,
- "amazon.titan-text-lite-v1": 4000,
- "amazon.titan-text-express-v1": 8000,
- "cohere.command-text-v14": 4000,
- "ai21.j2-mid-v1": 8191,
- "ai21.j2-ultra-v1": 8191,
- "ai21.jamba-instruct-v1:0": 256000,
- "mistral.mistral-7b-instruct-v0:2": 32000,
- "mistral.mixtral-8x7b-instruct-v0:1": 32000,
- "mistral-tiny": 32768,
- "mistral-small-latest": 32768,
- "mistral-medium-latest": 32768,
- "mistral-large-latest": 32768,
- "mistral-large-2407": 32768,
- "mistral-large-2402": 32768,
- "mistral/mistral-tiny": 32768,
- "mistral/mistral-small-latest": 32768,
- "mistral/mistral-medium-latest": 32768,
- "mistral/mistral-large-latest": 32768,
- "mistral/mistral-large-2407": 32768,
- "mistral/mistral-large-2402": 32768,
-}
-
-DEFAULT_CONTEXT_WINDOW_SIZE: Final[int] = 8192
-CONTEXT_WINDOW_USAGE_RATIO: Final[float] = 0.85
SUPPORTED_NATIVE_PROVIDERS: Final[list[str]] = [
"openai",
"anthropic",
@@ -2541,6 +2385,22 @@ class LLM(BaseLLM):
return remainder
return self.model
+ def _context_window_sizes(self) -> Mapping[str, int]:
+ """Pick a provider catalog when the model id is provider-qualified.
+
+ The merged catalog lets Azure's ``gpt-4`` override OpenAI's. Qualified
+ LiteLLM ids must keep the provider that was named.
+ """
+ prefix, separator, _remainder = self.model.partition("/")
+ if not separator:
+ return LLM_CONTEXT_WINDOW_SIZES
+ provider = prefix.lower()
+ if provider == "openai":
+ return OPENAI_CONTEXT_WINDOWS
+ if provider in {"azure", "azure_openai"}:
+ return AZURE_OPENAI_CONTEXT_WINDOWS
+ return LLM_CONTEXT_WINDOW_SIZES
+
def get_context_window_size(self) -> int:
"""
Returns the context window size, using 75% of the maximum to avoid
@@ -2552,22 +2412,12 @@ class LLM(BaseLLM):
if self.context_window_size != 0:
return self.context_window_size
- min_context = 1024
- max_context = 2097152 # Current max from gemini-1.5-pro
-
- for key, value in LLM_CONTEXT_WINDOW_SIZES.items():
- if value < min_context or value > max_context:
- raise ValueError(
- f"Context window for {key} must be between {min_context} and {max_context}"
- )
-
- self.context_window_size = int(
- DEFAULT_CONTEXT_WINDOW_SIZE * CONTEXT_WINDOW_USAGE_RATIO
+ self.context_window_size = resolve_context_window_size(
+ self._context_window_model_name(),
+ self._context_window_sizes(),
+ default=DEFAULT_CONTEXT_WINDOW_SIZE,
+ extra_names=(self.model,),
)
- model_name = self._context_window_model_name()
- for key, value in LLM_CONTEXT_WINDOW_SIZES.items():
- if model_name.startswith(key) or self.model.startswith(key):
- self.context_window_size = int(value * CONTEXT_WINDOW_USAGE_RATIO)
return self.context_window_size
@staticmethod
diff --git a/lib/crewai/src/crewai/llms/context_window.py b/lib/crewai/src/crewai/llms/context_window.py
new file mode 100644
index 000000000..92de8de9f
--- /dev/null
+++ b/lib/crewai/src/crewai/llms/context_window.py
@@ -0,0 +1,348 @@
+"""Shared context-window definitions and lookup for CrewAI LLMs."""
+
+from __future__ import annotations
+
+from collections.abc import Mapping, Sequence
+from typing import Final
+
+
+MIN_CONTEXT: Final[int] = 1024
+MAX_CONTEXT: Final[int] = 10_000_000
+DEFAULT_CONTEXT_WINDOW_SIZE: Final[int] = 8192
+CONTEXT_WINDOW_USAGE_RATIO: Final[float] = 0.85
+
+# Raw provider limits. ``resolve_context_window_size`` applies the usable
+# context ratio uniformly, so all call sites share one matching rule.
+# Keep active IDs and context windows in sync with OpenAI's model catalog;
+# remove or replace entries when the vendor marks them retired:
+# https://developers.openai.com/api/docs/models
+OPENAI_CONTEXT_WINDOWS: Final[dict[str, int]] = {
+ "gpt-4.1-mini-2025-04-14": 1_047_576,
+ "gpt-4.1-nano-2025-04-14": 1_047_576,
+ "gpt-6": 1_050_000,
+ "gpt-5.5": 1_050_000,
+ "gpt-5.4": 1_050_000,
+ "gpt-5.4-mini": 400_000,
+ "gpt-5.4-nano": 400_000,
+ "gpt-5.3-codex": 400_000,
+ "gpt-4-turbo": 128_000,
+ "gpt-4o-mini": 128_000,
+ "gpt-5-mini": 400_000,
+ "gpt-5-nano": 400_000,
+ "gpt-5.6": 1_050_000,
+ "gpt-5.6-cyber": 400_000,
+ "gpt-3.5-turbo": 16_385,
+ "gpt-oss-120b": 131_072,
+ "gpt-oss-20b": 131_072,
+ "chat-latest": 400_000,
+ "gpt-daybreak-blue-latest": 1_050_000,
+ "gpt-daybreak-red-latest": 400_000,
+ "o1-preview": 128_000,
+ "o1-mini": 128_000,
+ "o1-pro": 200_000,
+ "o1": 200_000,
+ "o3": 200_000,
+ "o3-mini": 200_000,
+ "o4-mini": 200_000,
+ "gpt-4.1": 1_047_576,
+ "gpt-4o": 128_000,
+ "gpt-5": 400_000,
+ "gpt-4": 8192,
+}
+
+# Check Azure deployment IDs and context windows here; availability varies by
+# account and region. Remove or replace vendor-retired entries:
+# https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
+AZURE_CONTEXT_WINDOWS: Final[dict[str, int]] = {
+ "gpt-6-astra": 1_050_000,
+ "gpt-5.6-sol": 1_050_000,
+ "gpt-5.6-terra": 1_050_000,
+ "gpt-5.6-luna": 1_050_000,
+ "gpt-chat-latest": 400_000,
+ "gpt-5.4-pro": 1_050_000,
+ "gpt-5.2-codex": 400_000,
+ "gpt-5.2": 400_000,
+ "gpt-5.1-codex-mini": 400_000,
+ "gpt-5.1-codex-max": 400_000,
+ "gpt-5.1-codex": 400_000,
+ "gpt-5.1": 400_000,
+ "gpt-5-codex": 400_000,
+ "gpt-5-pro": 400_000,
+ "gpt-oss-120b": 131_072,
+ "gpt-oss-20b": 131_072,
+ "codex-mini": 200_000,
+ "o3-pro": 200_000,
+ "computer-use-preview": 8192,
+ "gpt-4": 128_000,
+ "text-embedding-3-large": 8192,
+ "text-embedding-3-small": 8192,
+ "text-embedding-ada-002": 8192,
+ "text-embedding": 8191,
+}
+
+AZURE_OPENAI_CONTEXT_WINDOWS: Final[dict[str, int]] = {
+ **OPENAI_CONTEXT_WINDOWS,
+ **AZURE_CONTEXT_WINDOWS,
+}
+
+# Check active Claude IDs and context windows here; review retirements before
+# removing legacy prefixes: https://docs.anthropic.com/en/docs/about-claude/models
+# https://docs.anthropic.com/en/docs/about-claude/model-deprecations
+ANTHROPIC_CONTEXT_WINDOWS: Final[dict[str, int]] = {
+ "claude-fable-5-1": 1_000_000,
+ "claude-fable-5": 1_000_000,
+ "claude-mythos-5-1": 1_000_000,
+ "claude-mythos-5": 1_000_000,
+ "claude-opus-5-5": 1_000_000,
+ "claude-opus-5": 1_000_000,
+ "claude-sonnet-5": 1_000_000,
+ "claude-opus-4-8": 1_000_000,
+ "claude-opus-4-7": 1_000_000,
+ "claude-sonnet-4-6": 1_000_000,
+ "claude-opus-4-6": 1_000_000,
+ "claude-opus-4-5": 200_000,
+ "claude-sonnet-4-5": 200_000,
+ "claude-haiku-4-5": 200_000,
+}
+
+# Check Gemini model IDs and context windows here; remove or replace retired
+# entries listed in the catalog:
+# https://ai.google.dev/gemini-api/docs/models
+GEMINI_CONTEXT_WINDOWS: Final[dict[str, int]] = {
+ "gemini-3.8-flash": 1_048_576,
+ "gemini-3.7-flash": 1_048_576,
+ "gemini-3.6-flash": 1_048_576,
+ "gemini-3.5-flash": 1_048_576,
+ "gemini-3.5-flash-lite": 1_048_576,
+ "gemini-2.5-flash": 1_048_576,
+ "gemini-2.5-flash-lite": 1_048_576,
+ "gemini-2.5-pro": 1_048_576,
+ "gemini-3.1-flash-lite": 1_048_576,
+ "gemini-3.1-pro-preview": 1_048_576,
+ "gemini-3-flash-preview": 1_048_576,
+ "gemma-3-27b": 128_000,
+ "gemma-3-12b": 128_000,
+ "gemma-3-4b": 128_000,
+ "gemma-3-1b": 32_000,
+}
+
+# Check Bedrock model IDs and each model's context window in Models at a glance.
+# Availability varies by account and region; remove or replace retired entries:
+# https://docs.aws.amazon.com/bedrock/latest/userguide/models.html
+_BEDROCK_BASE_CONTEXT_WINDOWS: Final[dict[str, int]] = {
+ "amazon.nova-2-lite-v1:0": 1_000_000,
+ "amazon.nova-pro-v1:0": 300_000,
+ "amazon.nova-micro-v1:0": 128_000,
+ "amazon.nova-lite-v1:0": 300_000,
+ "ai21.jamba-1-5-large-v1:0": 256_000,
+ "ai21.jamba-1-5-mini-v1:0": 256_000,
+ "deepseek.r1-v1:0": 128_000,
+ "deepseek.v3-v1:0": 128_000,
+ "deepseek.v3.2": 164_000,
+ "google.gemma-3-12b-it": 128_000,
+ "google.gemma-3-27b-it": 128_000,
+ "google.gemma-3-4b-it": 128_000,
+ "google.gemma-4-26b-a4b": 256_000,
+ "google.gemma-4-31b": 256_000,
+ "google.gemma-4-e2b": 128_000,
+ "meta.llama3-1-405b-instruct-v1:0": 128_000,
+ "meta.llama3-3-70b-instruct-v1:0": 128_000,
+ "meta.llama3-1-70b-instruct-v1:0": 128_000,
+ "meta.llama3-1-8b-instruct-v1:0": 128_000,
+ "meta.llama3-70b-instruct-v1:0": 8000,
+ "meta.llama3-8b-instruct-v1:0": 8000,
+ "meta.llama3-2-11b-instruct-v1:0": 128_000,
+ "meta.llama3-2-3b-instruct-v1:0": 128_000,
+ "meta.llama3-2-90b-instruct-v1:0": 128_000,
+ "meta.llama3-2-1b-instruct-v1:0": 128_000,
+ "meta.llama4-maverick-17b-instruct-v1:0": 1_000_000,
+ "meta.llama4-scout-17b-instruct-v1:0": 10_000_000,
+ "minimax.minimax-m2": 1_000_000,
+ "minimax.minimax-m2.1": 196_000,
+ "minimax.minimax-m2.5": 196_000,
+ "mistral.devstral-2-123b": 256_000,
+ "mistral.magistral-small-2509": 128_000,
+ "mistral.ministral-3-14b-instruct": 128_000,
+ "mistral.ministral-3-8b-instruct": 128_000,
+ "mistral.ministral-3-3b-instruct": 128_000,
+ "mistral.mistral-7b-instruct-v0:2": 32_000,
+ "mistral.mistral-large-3-675b-instruct": 256_000,
+ "mistral.mistral-large-2402-v1:0": 32_000,
+ "mistral.mistral-small-2402-v1:0": 32_000,
+ "mistral.mixtral-8x7b-instruct-v0:1": 32_000,
+ "mistral.pixtral-large-2502-v1:0": 128_000,
+ "moonshot.kimi-k2-thinking": 256_000,
+ "moonshotai.kimi-k2.5": 256_000,
+ "moonshotai.kimi-k3": 1_000_000,
+ "nvidia.nemotron-nano-12b-v2": 128_000,
+ "nvidia.nemotron-nano-9b-v2": 128_000,
+ "nvidia.nemotron-nano-3-30b": 256_000,
+ "nvidia.nemotron-super-3-120b": 256_000,
+ "openai.gpt-5.4": 1_000_000,
+ "openai.gpt-5.5": 1_000_000,
+ "openai.gpt-5.6-cyber": 272_000,
+ "openai.gpt-5.6-luna": 1_000_000,
+ "openai.gpt-5.6-sol": 1_000_000,
+ "openai.gpt-5.6-terra": 1_000_000,
+ "openai.gpt-6-astra": 1_050_000,
+ "openai.gpt-6-luna": 1_050_000,
+ "openai.gpt-6-sol": 1_050_000,
+ "openai.gpt-daybreak-blue-5.6-sol": 1_000_000,
+ "openai.gpt-oss-20b-1:0": 128_000,
+ "openai.gpt-oss-120b-1:0": 128_000,
+ "openai.gpt-oss-safeguard-20b": 128_000,
+ "openai.gpt-oss-safeguard-120b": 128_000,
+ "qwen.qwen3-235b-a22b-2507-v1:0": 256_000,
+ "qwen.qwen3-32b-v1:0": 32_000,
+ "qwen.qwen3-coder-30b-a3b-v1:0": 256_000,
+ "qwen.qwen3-coder-480b-a35b-v1:0": 128_000,
+ "qwen.qwen3-coder-next": 256_000,
+ "qwen.qwen3-next-80b-a3b": 256_000,
+ "qwen.qwen3-vl-235b-a22b": 256_000,
+ "writer.palmyra-x4-v1:0": 128_000,
+ "writer.palmyra-x5-v1:0": 128_000,
+ "xai.grok-4.3": 1_000_000,
+ "xai.grok-4.6": 500_000,
+ "zai.glm-4.7": 203_000,
+ "zai.glm-4.7-flash": 203_000,
+ "zai.glm-5": 200_000,
+}
+
+
+def _prefixed_context_windows(
+ sizes: Mapping[str, int], prefixes: Sequence[str]
+) -> dict[str, int]:
+ return {
+ f"{prefix}{model}": size for prefix in prefixes for model, size in sizes.items()
+ }
+
+
+# Bedrock's lifecycle is independent from Anthropic's direct API. Sonnet 4 is
+# Legacy (not EOL) on Bedrock, so keep its provider-specific limit until AWS
+# retires it: https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle-legacy.html
+_BEDROCK_LEGACY_CONTEXT_WINDOWS: Final[dict[str, int]] = {
+ "anthropic.claude-sonnet-4": 200_000,
+ "us.anthropic.claude-sonnet-4": 200_000,
+ "eu.anthropic.claude-sonnet-4": 200_000,
+ "apac.anthropic.claude-sonnet-4": 200_000,
+ "global.anthropic.claude-sonnet-4": 200_000,
+ **_prefixed_context_windows(
+ {
+ "anthropic.claude-opus-4-1": 200_000,
+ "anthropic.claude-opus-4": 200_000,
+ "anthropic.claude-haiku-4": 200_000,
+ "anthropic.claude-3-7-sonnet": 200_000,
+ "anthropic.claude-3-5-sonnet": 200_000,
+ "anthropic.claude-3-5-haiku": 200_000,
+ "anthropic.claude-3-opus": 200_000,
+ "anthropic.claude-3-sonnet": 200_000,
+ "anthropic.claude-3-haiku": 200_000,
+ },
+ ("", "us.", "eu.", "apac.", "global."),
+ ),
+}
+
+# LiteLLM maintains its provider-only model IDs and context windows here. Prefer
+# the vendor catalogs above when both sources define the same model:
+# https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json
+LITELLM_CONTEXT_WINDOWS: Final[dict[str, int]] = {
+ "deepseek-chat": 128_000,
+ "claude-3-5-sonnet": 200_000,
+ "gemma2-9b-it": 8192,
+ "gemma-7b-it": 8192,
+ "llama3-groq-70b-8192-tool-use-preview": 8192,
+ "llama3-groq-8b-8192-tool-use-preview": 8192,
+ "llama-3.1-70b-versatile": 131_072,
+ "llama-3.1-8b-instant": 131_072,
+ "llama-3.2-1b-preview": 8192,
+ "llama-3.2-3b-preview": 8192,
+ "llama-3.2-11b-text-preview": 8192,
+ "llama-3.2-90b-text-preview": 8192,
+ "llama3-70b-8192": 8192,
+ "llama3-8b-8192": 8192,
+ "mixtral-8x7b-32768": 32_768,
+ "llama-3.3-70b-versatile": 128_000,
+ "llama-3.3-70b-instruct": 128_000,
+ "Meta-Llama-3.3-70B-Instruct": 131_072,
+ "QwQ-32B-Preview": 8192,
+ "Qwen2.5-72B-Instruct": 8192,
+ "Qwen2.5-Coder-32B-Instruct": 8192,
+ "Meta-Llama-3.1-405B-Instruct": 8192,
+ "Meta-Llama-3.1-70B-Instruct": 131_072,
+ "Meta-Llama-3.1-8B-Instruct": 131_072,
+ "Llama-3.2-90B-Vision-Instruct": 16_384,
+ "Llama-3.2-11B-Vision-Instruct": 16_384,
+ "Meta-Llama-3.2-3B-Instruct": 4096,
+ "Meta-Llama-3.2-1B-Instruct": 16_384,
+ "gemini/gemma-3-1b-it": 32_000,
+ "gemini/gemma-3-4b-it": 128_000,
+ "gemini/gemma-3-12b-it": 128_000,
+ "gemini/gemma-3-27b-it": 128_000,
+ "mistral-tiny": 32_768,
+ "mistral-small-latest": 32_768,
+ "mistral-medium-latest": 32_768,
+ "mistral-large-latest": 32_768,
+ "mistral-large-2407": 32_768,
+ "mistral-large-2402": 32_768,
+ "mistral/mistral-tiny": 32_768,
+ "mistral/mistral-small-latest": 32_768,
+ "mistral/mistral-medium-latest": 32_768,
+ "mistral/mistral-large-latest": 32_768,
+ "mistral/mistral-large-2407": 32_768,
+ "mistral/mistral-large-2402": 32_768,
+}
+
+
+BEDROCK_CONTEXT_WINDOWS: Final[dict[str, int]] = {
+ **_BEDROCK_BASE_CONTEXT_WINDOWS,
+ **_BEDROCK_LEGACY_CONTEXT_WINDOWS,
+ **_prefixed_context_windows(_BEDROCK_BASE_CONTEXT_WINDOWS, ("us.", "eu.", "apac.")),
+ **_prefixed_context_windows(
+ ANTHROPIC_CONTEXT_WINDOWS,
+ (
+ "anthropic.",
+ "us.anthropic.",
+ "eu.anthropic.",
+ "apac.anthropic.",
+ "global.anthropic.",
+ ),
+ ),
+ # Geo inference ID is not produced by the us/eu/apac expansion.
+ "global.moonshotai.kimi-k3": 1_000_000,
+}
+
+LLM_CONTEXT_WINDOW_SIZES: Final[dict[str, int]] = {
+ **LITELLM_CONTEXT_WINDOWS,
+ **OPENAI_CONTEXT_WINDOWS,
+ **AZURE_CONTEXT_WINDOWS,
+ **ANTHROPIC_CONTEXT_WINDOWS,
+ **GEMINI_CONTEXT_WINDOWS,
+ **BEDROCK_CONTEXT_WINDOWS,
+}
+
+
+def resolve_context_window_size(
+ model: str,
+ sizes: Mapping[str, int],
+ *,
+ default: int,
+ extra_names: Sequence[str] = (),
+) -> int:
+ """Return the usable context window for a model using its longest prefix."""
+ for name, size in sizes.items():
+ if size < MIN_CONTEXT or size > MAX_CONTEXT:
+ raise ValueError(
+ f"Context window for {name} must be between {MIN_CONTEXT} and {MAX_CONTEXT}"
+ )
+
+ candidates = (model, *extra_names)
+ _, size = max(
+ (
+ (prefix, size)
+ for prefix, size in sizes.items()
+ if any(name.startswith(prefix) for name in candidates)
+ ),
+ key=lambda match: len(match[0]),
+ default=("", default),
+ )
+ return int(size * CONTEXT_WINDOW_USAGE_RATIO)
diff --git a/lib/crewai/src/crewai/llms/providers/anthropic/completion.py b/lib/crewai/src/crewai/llms/providers/anthropic/completion.py
index 77236249a..366663ccc 100644
--- a/lib/crewai/src/crewai/llms/providers/anthropic/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/anthropic/completion.py
@@ -1972,28 +1972,14 @@ class AnthropicCompletion(BaseLLM):
def get_context_window_size(self) -> int:
"""Get the context window size for the model."""
- from crewai.llm import CONTEXT_WINDOW_USAGE_RATIO
-
- # Current offered models. Unknown and retired IDs fall back to 200k.
- context_windows = {
- "claude-fable-5": 1000000,
- "claude-mythos-5": 1000000,
- "claude-opus-5": 1000000,
- "claude-sonnet-5": 1000000,
- "claude-opus-4-8": 1000000,
- "claude-opus-4-7": 1000000,
- "claude-opus-4-6": 1000000,
- "claude-sonnet-4-6": 1000000,
- "claude-opus-4-5": 200000,
- "claude-sonnet-4-5": 200000,
- "claude-haiku-4-5": 200000,
- }
-
- for model_prefix, size in context_windows.items():
- if self.model.startswith(model_prefix):
- return int(size * CONTEXT_WINDOW_USAGE_RATIO)
+ from crewai.llms.context_window import (
+ ANTHROPIC_CONTEXT_WINDOWS,
+ resolve_context_window_size,
+ )
- return int(200000 * CONTEXT_WINDOW_USAGE_RATIO)
+ return resolve_context_window_size(
+ self.model, ANTHROPIC_CONTEXT_WINDOWS, default=200_000
+ )
@staticmethod
def _extract_finish_reason_and_id(
diff --git a/lib/crewai/src/crewai/llms/providers/azure/completion.py b/lib/crewai/src/crewai/llms/providers/azure/completion.py
index 0f45bfcef..7b7f7d9f0 100644
--- a/lib/crewai/src/crewai/llms/providers/azure/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/azure/completion.py
@@ -1308,36 +1308,17 @@ class AzureCompletion(BaseLLM):
def get_context_window_size(self) -> int:
"""Get the context window size for the model."""
- from crewai.llm import CONTEXT_WINDOW_USAGE_RATIO, LLM_CONTEXT_WINDOW_SIZES
-
- min_context = 1024
- max_context = 2097152
-
- for key, value in LLM_CONTEXT_WINDOW_SIZES.items():
- if value < min_context or value > max_context:
- raise ValueError(
- f"Context window for {key} must be between {min_context} and {max_context}"
- )
-
- # Longest prefix first. Always insert new keys in that order so
- # startswith prefers gpt-5.6 over gpt-5, gpt-4o-mini over gpt-4o, etc.
- context_windows = {
- "text-embedding": 8191,
- "gpt-3.5-turbo": 16385,
- "gpt-5.4-mini": 200000,
- "gpt-35-turbo": 16385,
- "gpt-4o-mini": 128000,
- "gpt-4-turbo": 128000,
- "gpt-5.6": 1050000,
- "gpt-4o": 128000,
- "gpt-4": 8192,
- }
-
- for model_prefix, size in context_windows.items():
- if self.model.startswith(model_prefix):
- return int(size * CONTEXT_WINDOW_USAGE_RATIO)
+ from crewai.llms.context_window import (
+ AZURE_OPENAI_CONTEXT_WINDOWS,
+ DEFAULT_CONTEXT_WINDOW_SIZE,
+ resolve_context_window_size,
+ )
- return int(8192 * CONTEXT_WINDOW_USAGE_RATIO)
+ return resolve_context_window_size(
+ self.model,
+ AZURE_OPENAI_CONTEXT_WINDOWS,
+ default=DEFAULT_CONTEXT_WINDOW_SIZE,
+ )
def _effective_max_tokens(self) -> int | float | None:
"""Azure reasoning/newer chat models cap via ``max_completion_tokens``."""
diff --git a/lib/crewai/src/crewai/llms/providers/bedrock/completion.py b/lib/crewai/src/crewai/llms/providers/bedrock/completion.py
index 2528aa204..cf0d552ae 100644
--- a/lib/crewai/src/crewai/llms/providers/bedrock/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/bedrock/completion.py
@@ -2135,33 +2135,15 @@ class BedrockCompletion(BaseLLM):
def get_context_window_size(self) -> int:
"""Get the context window size for the model."""
- from crewai.llm import CONTEXT_WINDOW_USAGE_RATIO
-
- context_windows = {
- "anthropic.claude-sonnet-4": 200000,
- "anthropic.claude-opus-4": 200000,
- "anthropic.claude-haiku-4": 200000,
- "anthropic.claude-3-5-sonnet": 200000,
- "anthropic.claude-3-5-haiku": 200000,
- "anthropic.claude-3-opus": 200000,
- "anthropic.claude-3-sonnet": 200000,
- "anthropic.claude-3-haiku": 200000,
- "anthropic.claude-3-7-sonnet": 200000,
- "anthropic.claude-v2": 100000,
- "amazon.titan-text-express": 8000,
- "ai21.j2-ultra": 8192,
- "cohere.command-text": 4096,
- "meta.llama2-13b-chat": 4096,
- "meta.llama2-70b-chat": 4096,
- "meta.llama3-70b-instruct": 128000,
- "deepseek.r1": 32768,
- }
-
- for model_prefix, size in context_windows.items():
- if self.model.startswith(model_prefix):
- return int(size * CONTEXT_WINDOW_USAGE_RATIO)
+ from crewai.llms.context_window import (
+ BEDROCK_CONTEXT_WINDOWS,
+ DEFAULT_CONTEXT_WINDOW_SIZE,
+ resolve_context_window_size,
+ )
- return int(8192 * CONTEXT_WINDOW_USAGE_RATIO)
+ return resolve_context_window_size(
+ self.model, BEDROCK_CONTEXT_WINDOWS, default=DEFAULT_CONTEXT_WINDOW_SIZE
+ )
def supports_multimodal(self) -> bool:
"""Check if the model supports multimodal inputs.
diff --git a/lib/crewai/src/crewai/llms/providers/gemini/completion.py b/lib/crewai/src/crewai/llms/providers/gemini/completion.py
index 6ca6da229..1129d4112 100644
--- a/lib/crewai/src/crewai/llms/providers/gemini/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/gemini/completion.py
@@ -1382,40 +1382,14 @@ class GeminiCompletion(BaseLLM):
def get_context_window_size(self) -> int:
"""Get the context window size for the model."""
- from crewai.llm import CONTEXT_WINDOW_USAGE_RATIO, LLM_CONTEXT_WINDOW_SIZES
-
- min_context = 1024
- max_context = 2097152
-
- for key, value in LLM_CONTEXT_WINDOW_SIZES.items():
- if value < min_context or value > max_context:
- raise ValueError(
- f"Context window for {key} must be between {min_context} and {max_context}"
- )
-
- context_windows = {
- "gemini-3.8-flash": 1048576, # 1M tokens
- "gemini-3-pro-preview": 1048576, # 1M tokens
- "gemini-2.0-flash": 1048576, # 1M tokens
- "gemini-2.0-flash-thinking": 32768,
- "gemini-2.0-flash-lite": 1048576,
- "gemini-2.5-flash": 1048576,
- "gemini-2.5-pro": 1048576,
- "gemini-1.5-pro": 2097152, # 2M tokens
- "gemini-1.5-flash": 1048576,
- "gemini-1.5-flash-8b": 1048576,
- "gemini-1.0-pro": 32768,
- "gemma-3-1b": 32000,
- "gemma-3-4b": 128000,
- "gemma-3-12b": 128000,
- "gemma-3-27b": 128000,
- }
-
- for model_prefix, size in context_windows.items():
- if self.model.startswith(model_prefix):
- return int(size * CONTEXT_WINDOW_USAGE_RATIO)
+ from crewai.llms.context_window import (
+ GEMINI_CONTEXT_WINDOWS,
+ resolve_context_window_size,
+ )
- return int(1048576 * CONTEXT_WINDOW_USAGE_RATIO) # 1M tokens default
+ return resolve_context_window_size(
+ self.model, GEMINI_CONTEXT_WINDOWS, default=1_048_576
+ )
def _effective_max_tokens(self) -> int | float | None:
"""Gemini caps generation via ``max_output_tokens``."""
diff --git a/lib/crewai/src/crewai/llms/providers/openai/completion.py b/lib/crewai/src/crewai/llms/providers/openai/completion.py
index 88164ba64..26d8fc7ab 100644
--- a/lib/crewai/src/crewai/llms/providers/openai/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/openai/completion.py
@@ -2968,43 +2968,15 @@ class OpenAICompletion(BaseLLM):
def get_context_window_size(self) -> int:
"""Get the context window size for the model."""
- from crewai.llm import CONTEXT_WINDOW_USAGE_RATIO, LLM_CONTEXT_WINDOW_SIZES
-
- min_context = 1024
- max_context = 2097152
-
- for key, value in LLM_CONTEXT_WINDOW_SIZES.items():
- if value < min_context or value > max_context:
- raise ValueError(
- f"Context window for {key} must be between {min_context} and {max_context}"
- )
-
- # Longest prefix first. Always insert new keys in that order so
- # startswith prefers gpt-5.6 over gpt-5, gpt-4o-mini over gpt-4o, etc.
- context_windows = {
- "gpt-4.1-mini-2025-04-14": 1047576,
- "gpt-4.1-nano-2025-04-14": 1047576,
- "gpt-5.4-mini": 200000,
- "gpt-4-turbo": 128000,
- "gpt-4o-mini": 128000,
- "gpt-5-mini": 1047576,
- "gpt-5-nano": 1047576,
- "o1-preview": 128000,
- "gpt-5.6": 1050000,
- "o1-mini": 128000,
- "o3-mini": 200000,
- "o4-mini": 200000,
- "gpt-4.1": 1047576,
- "gpt-4o": 128000,
- "gpt-5": 1047576,
- "gpt-4": 8192,
- }
-
- for model_prefix, size in context_windows.items():
- if self.model.startswith(model_prefix):
- return int(size * CONTEXT_WINDOW_USAGE_RATIO)
+ from crewai.llms.context_window import (
+ DEFAULT_CONTEXT_WINDOW_SIZE,
+ OPENAI_CONTEXT_WINDOWS,
+ resolve_context_window_size,
+ )
- return int(8192 * CONTEXT_WINDOW_USAGE_RATIO)
+ return resolve_context_window_size(
+ self.model, OPENAI_CONTEXT_WINDOWS, default=DEFAULT_CONTEXT_WINDOW_SIZE
+ )
def _effective_max_tokens(self) -> int | float | None:
"""Newer OpenAI chat models cap via ``max_completion_tokens``."""
diff --git a/lib/crewai/tests/llms/anthropic/test_anthropic.py b/lib/crewai/tests/llms/anthropic/test_anthropic.py
index 1308fdad0..5056393ce 100644
--- a/lib/crewai/tests/llms/anthropic/test_anthropic.py
+++ b/lib/crewai/tests/llms/anthropic/test_anthropic.py
@@ -5,7 +5,8 @@ from unittest.mock import AsyncMock, patch, MagicMock
import pytest
from pydantic import BaseModel
-from crewai.llm import CONTEXT_WINDOW_USAGE_RATIO, LLM
+from crewai.llm import LLM
+from crewai.llms.context_window import CONTEXT_WINDOW_USAGE_RATIO
from crewai.crew import Crew
from crewai.agent import Agent
from crewai.task import Task
diff --git a/lib/crewai/tests/llms/azure/test_azure.py b/lib/crewai/tests/llms/azure/test_azure.py
index 746d719aa..e30ae9294 100644
--- a/lib/crewai/tests/llms/azure/test_azure.py
+++ b/lib/crewai/tests/llms/azure/test_azure.py
@@ -5,7 +5,8 @@ from unittest.mock import patch, MagicMock, Mock
from urllib.parse import urlparse
import pytest
-from crewai.llm import CONTEXT_WINDOW_USAGE_RATIO, LLM
+from crewai.llm import LLM
+from crewai.llms.context_window import CONTEXT_WINDOW_USAGE_RATIO
from crewai.crew import Crew
from crewai.agent import Agent
from crewai.task import Task
@@ -644,13 +645,33 @@ def test_azure_context_window_size():
"""
Test that Azure models return correct context window sizes
"""
+ usable = int(128_000 * CONTEXT_WINDOW_USAGE_RATIO)
+
llm_gpt4 = LLM(model="azure/gpt-4")
- context_size_gpt4 = llm_gpt4.get_context_window_size()
- assert context_size_gpt4 > 0
+ assert llm_gpt4.get_context_window_size() == usable
llm_gpt4o = LLM(model="azure/gpt-4o")
- context_size_gpt4o = llm_gpt4o.get_context_window_size()
- assert context_size_gpt4o > context_size_gpt4 # GPT-4o has larger context
+ assert llm_gpt4o.get_context_window_size() == usable
+
+
+ ("model", "context_window"),
+ [
+ ("azure/gpt-6-astra", 1_050_000),
+ ("azure/gpt-chat-latest", 400_000),
+ ("azure/gpt-oss-120b", 131_072),
+ ("azure/codex-mini", 200_000),
+ ("azure/computer-use-preview", 8192),
+ ("azure/gpt-4", 128_000),
+ ],
+)
+def test_azure_specific_models_use_official_context_windows(
+ model: str, context_window: int
+) -> None:
+ llm = LLM(model=model)
+ assert llm.get_context_window_size() == int(
+ context_window * CONTEXT_WINDOW_USAGE_RATIO
+ )
@pytest.mark.parametrize(
@@ -670,7 +691,7 @@ def test_azure_gpt56_family_uses_official_context_window(model: str) -> None:
def test_azure_gpt54_mini_keeps_its_window() -> None:
llm = LLM(model="azure/gpt-5.4-mini")
- assert llm.get_context_window_size() == int(200000 * CONTEXT_WINDOW_USAGE_RATIO)
+ assert llm.get_context_window_size() == int(400_000 * CONTEXT_WINDOW_USAGE_RATIO)
def test_azure_message_formatting():
diff --git a/lib/crewai/tests/llms/bedrock/test_bedrock.py b/lib/crewai/tests/llms/bedrock/test_bedrock.py
index 99eccdd1e..6d9b67cdf 100644
--- a/lib/crewai/tests/llms/bedrock/test_bedrock.py
+++ b/lib/crewai/tests/llms/bedrock/test_bedrock.py
@@ -5,6 +5,7 @@ from unittest.mock import AsyncMock, MagicMock, patch
import pytest
from crewai.llm import LLM
+from crewai.llms.context_window import CONTEXT_WINDOW_USAGE_RATIO
from crewai.crew import Crew
from crewai.agent import Agent
from crewai.task import Task
@@ -543,15 +544,23 @@ def test_bedrock_context_window_size():
"""
Test that Bedrock models return correct context window sizes
"""
- llm_claude = LLM(model="bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0")
- context_size_claude = llm_claude.get_context_window_size()
- assert context_size_claude > 150000 # Should be substantial (200K tokens with ratio)
+ llm_nova = LLM(model="bedrock/amazon.nova-2-lite-v1:0")
+ context_size_nova = llm_nova.get_context_window_size()
+ assert context_size_nova == int(1_000_000 * CONTEXT_WINDOW_USAGE_RATIO)
llm_titan = LLM(model="bedrock/amazon.titan-text-express-v1")
context_size_titan = llm_titan.get_context_window_size()
assert context_size_titan > 5000
+def test_bedrock_claude_sonnet_46_uses_its_specific_context_window():
+ llm = LLM(model="bedrock/anthropic.claude-sonnet-4-6-v1:0")
+
+ assert llm.get_context_window_size() == int(
+ 1_000_000 * CONTEXT_WINDOW_USAGE_RATIO
+ )
+
+
def test_bedrock_message_formatting():
"""
Test that messages are properly formatted for Bedrock Converse API
diff --git a/lib/crewai/tests/llms/google/test_google.py b/lib/crewai/tests/llms/google/test_google.py
index 31fe5f8fa..af9095601 100644
--- a/lib/crewai/tests/llms/google/test_google.py
+++ b/lib/crewai/tests/llms/google/test_google.py
@@ -485,20 +485,20 @@ def test_gemini_context_window_size():
"""
Test that Gemini models return correct context window sizes
"""
- # Test Gemini 2.0 Flash
- llm_2_0 = LLM(model="google/gemini-2.0-flash-001")
- context_size_2_0 = llm_2_0.get_context_window_size()
- assert context_size_2_0 > 500000
+ # Test Gemini 2.5 Flash
+ llm_2_5 = LLM(model="google/gemini-2.5-flash")
+ context_size_2_5 = llm_2_5.get_context_window_size()
+ assert context_size_2_5 > 500000
# Test Gemini 3.8 Flash
llm_3_8 = LLM(model="google/gemini-3.8-flash")
context_size_3_8 = llm_3_8.get_context_window_size()
assert context_size_3_8 == 891289
- # Test Gemini 1.5 Pro
- llm_1_5 = LLM(model="google/gemini-1.5-pro")
- context_size_1_5 = llm_1_5.get_context_window_size()
- assert context_size_1_5 > 1000000
+ # Test Gemini 2.5 Pro
+ llm_2_5_pro = LLM(model="google/gemini-2.5-pro")
+ context_size_2_5_pro = llm_2_5_pro.get_context_window_size()
+ assert context_size_2_5_pro > 500000
def test_gemini_message_formatting():
diff --git a/lib/crewai/tests/llms/openai/test_openai.py b/lib/crewai/tests/llms/openai/test_openai.py
index 994092acd..5a06f4676 100644
--- a/lib/crewai/tests/llms/openai/test_openai.py
+++ b/lib/crewai/tests/llms/openai/test_openai.py
@@ -7,7 +7,8 @@ from unittest.mock import patch, MagicMock
import openai
import pytest
-from crewai.llm import CONTEXT_WINDOW_USAGE_RATIO, LLM
+from crewai.llm import LLM
+from crewai.llms.context_window import CONTEXT_WINDOW_USAGE_RATIO
from crewai.llms.providers.openai.completion import OpenAICompletion, ResponsesAPIResult
from crewai.crew import Crew
from crewai.agent import Agent
@@ -1877,8 +1878,8 @@ def test_openai_prefixed_gpt56_luna_uses_official_context_window() -> None:
def test_openai_gpt5_and_gpt54_mini_keep_their_windows() -> None:
gpt5 = OpenAICompletion(model="gpt-5")
gpt54_mini = OpenAICompletion(model="gpt-5.4-mini")
- assert gpt5.get_context_window_size() == int(1_047_576 * CONTEXT_WINDOW_USAGE_RATIO)
- assert gpt54_mini.get_context_window_size() == int(200000 * CONTEXT_WINDOW_USAGE_RATIO)
+ assert gpt5.get_context_window_size() == int(400_000 * CONTEXT_WINDOW_USAGE_RATIO)
+ assert gpt54_mini.get_context_window_size() == int(400_000 * CONTEXT_WINDOW_USAGE_RATIO)
def test_openai_stop_words_still_applied_to_regular_responses():
diff --git a/lib/crewai/tests/llms/test_context_window.py b/lib/crewai/tests/llms/test_context_window.py
new file mode 100644
index 000000000..d8d414764
--- /dev/null
+++ b/lib/crewai/tests/llms/test_context_window.py
@@ -0,0 +1,116 @@
+import pytest
+
+from crewai.llms.context_window import (
+ BEDROCK_CONTEXT_WINDOWS,
+ GEMINI_CONTEXT_WINDOWS,
+ CONTEXT_WINDOW_USAGE_RATIO,
+ DEFAULT_CONTEXT_WINDOW_SIZE,
+ MAX_CONTEXT,
+ LLM_CONTEXT_WINDOW_SIZES,
+ resolve_context_window_size,
+)
+
+
+def test_resolver_prefers_the_longest_matching_prefix() -> None:
+ sizes = {"gpt-5": 1_047_576, "gpt-5.4-mini": 200_000}
+
+ result = resolve_context_window_size(
+ "gpt-5.4-mini-2026-01-01", sizes, default=DEFAULT_CONTEXT_WINDOW_SIZE
+ )
+
+ assert result == int(200_000 * CONTEXT_WINDOW_USAGE_RATIO)
+
+
+def test_resolver_uses_default_for_an_unknown_model() -> None:
+ result = resolve_context_window_size(
+ "unknown-model", {}, default=DEFAULT_CONTEXT_WINDOW_SIZE
+ )
+
+ assert result == int(DEFAULT_CONTEXT_WINDOW_SIZE * CONTEXT_WINDOW_USAGE_RATIO)
+
+
+def test_resolver_validates_all_declared_context_windows() -> None:
+ with pytest.raises(ValueError, match="must be between 1024 and 10000000"):
+ resolve_context_window_size("test-model", {"test-model": 500}, default=8192)
+
+
+def test_bedrock_claude_aliases_use_the_anthropic_context_window() -> None:
+ result = resolve_context_window_size(
+ "us.anthropic.claude-sonnet-4-6-v1:0",
+ BEDROCK_CONTEXT_WINDOWS,
+ default=DEFAULT_CONTEXT_WINDOW_SIZE,
+ )
+
+ assert result == int(1_000_000 * CONTEXT_WINDOW_USAGE_RATIO)
+
+
+def test_bedrock_regional_aliases_preserve_the_base_model_context_window() -> None:
+ result = resolve_context_window_size(
+ "us.meta.llama3-3-70b-instruct-v1:0",
+ BEDROCK_CONTEXT_WINDOWS,
+ default=DEFAULT_CONTEXT_WINDOW_SIZE,
+ )
+
+ assert result == int(128_000 * CONTEXT_WINDOW_USAGE_RATIO)
+
+
+def test_litellm_map_includes_the_openai_gpt5_family() -> None:
+ result = resolve_context_window_size(
+ "gpt-5", LLM_CONTEXT_WINDOW_SIZES, default=DEFAULT_CONTEXT_WINDOW_SIZE
+ )
+
+ assert result == int(400_000 * CONTEXT_WINDOW_USAGE_RATIO)
+
+
+ ("model", "raw_context_window"),
+ [
+ ("o1", 200_000),
+ ("o1-pro", 200_000),
+ ("o3", 200_000),
+ ("gemini-3.1-flash-lite", 1_048_576),
+ ("gemini-3.1-pro-preview", 1_048_576),
+ ("gemini-3-flash-preview", 1_048_576),
+ ],
+)
+def test_affected_model_ids_use_their_specific_context_windows(
+ model: str, raw_context_window: int
+) -> None:
+ assert resolve_context_window_size(
+ model, LLM_CONTEXT_WINDOW_SIZES, default=DEFAULT_CONTEXT_WINDOW_SIZE
+ ) == int(raw_context_window * CONTEXT_WINDOW_USAGE_RATIO)
+
+
+ ("model", "sizes", "raw_context_window"),
+ [
+ ("gpt-6-astra", LLM_CONTEXT_WINDOW_SIZES, 1_050_000),
+ ("gpt-5.6-cyber", LLM_CONTEXT_WINDOW_SIZES, 400_000),
+ ("gpt-3.5-turbo", LLM_CONTEXT_WINDOW_SIZES, 16_385),
+ ("chat-latest", LLM_CONTEXT_WINDOW_SIZES, 400_000),
+ ("gpt-oss-120b", LLM_CONTEXT_WINDOW_SIZES, 131_072),
+ ("gpt-5.4-nano", LLM_CONTEXT_WINDOW_SIZES, 400_000),
+ ("gemini-3.7-flash", LLM_CONTEXT_WINDOW_SIZES, 1_048_576),
+ ("gemma-3-1b-it", GEMINI_CONTEXT_WINDOWS, 32_000),
+ ("gemma-3-27b-it", GEMINI_CONTEXT_WINDOWS, 128_000),
+ ("amazon.nova-2-lite-v1:0", BEDROCK_CONTEXT_WINDOWS, 1_000_000),
+ ("global.moonshotai.kimi-k3", BEDROCK_CONTEXT_WINDOWS, 1_000_000),
+ (
+ "anthropic.claude-3-5-sonnet-20241022-v2:0",
+ BEDROCK_CONTEXT_WINDOWS,
+ 200_000,
+ ),
+ ("us.anthropic.claude-opus-4-1-20250805-v1:0", BEDROCK_CONTEXT_WINDOWS, 200_000),
+ (
+ "meta.llama4-scout-17b-instruct-v1:0",
+ BEDROCK_CONTEXT_WINDOWS,
+ MAX_CONTEXT,
+ ),
+ ],
+)
+def test_new_catalog_models_use_their_documented_context_windows(
+ model: str, sizes: dict[str, int], raw_context_window: int
+) -> None:
+ assert resolve_context_window_size(
+ model, sizes, default=DEFAULT_CONTEXT_WINDOW_SIZE
+ ) == int(raw_context_window * CONTEXT_WINDOW_USAGE_RATIO)
diff --git a/lib/crewai/tests/test_llm.py b/lib/crewai/tests/test_llm.py
index 870b3f9e5..8cd1aab4d 100644
--- a/lib/crewai/tests/test_llm.py
+++ b/lib/crewai/tests/test_llm.py
@@ -11,7 +11,8 @@ from crewai.events.event_types import (
ToolUsageFinishedEvent,
ToolUsageStartedEvent,
)
-from crewai.llm import CONTEXT_WINDOW_USAGE_RATIO, DEFAULT_CONTEXT_WINDOW_SIZE, LLM
+from crewai.llm import DEFAULT_CONTEXT_WINDOW_SIZE, LLM
+from crewai.llms.context_window import CONTEXT_WINDOW_USAGE_RATIO
from crewai.llms.providers.anthropic.completion import AnthropicCompletion
from crewai.utilities.agent_utils import is_context_length_exceeded
from crewai.utilities.exceptions.context_window_exceeding_exception import (
@@ -343,9 +344,9 @@ def test_context_window_validation():
{"test-model": 500},
clear=True,
):
- llm = LLM(model="test-model")
+ llm = LLM(model="test-model", is_litellm=True)
llm.get_context_window_size()
- assert "must be between 1024 and 2097152" in str(excinfo.value)
+ assert "must be between 1024 and 10000000" in str(excinfo.value)
@pytest.mark.parametrize(
@@ -361,7 +362,43 @@ def test_gpt56_family_uses_official_context_window(model: str) -> None:
def test_gpt56_does_not_override_gpt54_mini_window() -> None:
"""A more specific older prefix must keep its own window."""
llm = LLM(model="gpt-5.4-mini", is_litellm=True)
- assert llm.get_context_window_size() == int(200000 * CONTEXT_WINDOW_USAGE_RATIO)
+ assert llm.get_context_window_size() == int(400000 * CONTEXT_WINDOW_USAGE_RATIO)
+
+
+def test_qualified_litellm_gpt4_keeps_the_provider_context_window() -> None:
+ openai_llm = LLM(model="openai/gpt-4", is_litellm=True)
+ azure_llm = LLM(model="azure/gpt-4", is_litellm=True)
+
+ assert openai_llm.get_context_window_size() == int(8192 * CONTEXT_WINDOW_USAGE_RATIO)
+ assert azure_llm.get_context_window_size() == int(
+ 128_000 * CONTEXT_WINDOW_USAGE_RATIO
+ )
+
+
+def test_litellm_gpt5_uses_the_shared_openai_context_window() -> None:
+ llm = LLM(model="gpt-5", is_litellm=True)
+
+ assert llm.get_context_window_size() == int(
+ 400_000 * CONTEXT_WINDOW_USAGE_RATIO
+ )
+
+
+ ("native_model", "litellm_model"),
+ [
+ ("openai/gpt-5", "gpt-5"),
+ ("openai/o1-pro", "o1-pro"),
+ ("anthropic/claude-sonnet-4-6", "claude-sonnet-4-6"),
+ ("google/gemini-2.5-flash", "gemini-2.5-flash"),
+ ],
+)
+def test_native_and_litellm_context_windows_match(
+ native_model: str, litellm_model: str
+) -> None:
+ native_llm = LLM(model=native_model)
+ litellm_llm = LLM(model=litellm_model, is_litellm=True)
+
+ assert native_llm.get_context_window_size() == litellm_llm.get_context_window_size()
@pytest.mark.parametrize(
diff --git a/lib/crewai/tests/utilities/test_agent_utils.py b/lib/crewai/tests/utilities/test_agent_utils.py
index a806eef75..9416d5dfd 100644
--- a/lib/crewai/tests/utilities/test_agent_utils.py
+++ b/lib/crewai/tests/utilities/test_agent_utils.py
@@ -18,7 +18,7 @@ from crewai.hooks.tool_hooks import (
)
from crewai.agents.parser import AgentFinish
from crewai.tools.base_tool import BaseTool
-from crewai.llm import CONTEXT_WINDOW_USAGE_RATIO
+from crewai.llms.context_window import CONTEXT_WINDOW_USAGE_RATIO
from crewai.utilities.agent_utils import (
_asummarize_chunks,
_estimate_token_count,