langfun 0.0.2.dev20240330__py3-none-any.whl → 0.1.2.dev202501140804__py3-none-any.whl

langfun/core/llms/openai.py

@@ -13,95 +13,302 @@
 # limitations under the License.
 """Language models from OpenAI."""
 
-import collections
-import functools
 import os
-from typing import Annotated, Any, cast
+from typing import Annotated, Any
 
 import langfun.core as lf
-from langfun.core import modalities as lf_modalities
-import openai
-from openai import error as openai_error
-from openai import openai_object
+from langfun.core.llms import openai_compatible
 import pyglove as pg
 
 
-class Usage(pg.Object):
-  """Usage information per completion."""
-
-  prompt_tokens: int
-  completion_tokens: int
-  total_tokens: int
-
-
-class LMSamplingResult(lf.LMSamplingResult):
-  """LMSamplingResult with usage information."""
-
-  usage: Usage | None = None
-
-
-SUPPORTED_MODELS_AND_SETTINGS = [
-    # Model name, max concurrent requests.
-    # The concurrent requests is estimated by TPM/RPM from
-    # https://platform.openai.com/account/limits
-    # GPT-4 Turbo models.
-    ('gpt-4-turbo-preview', 1),  # GPT-4 Turbo.
-    ('gpt-4-0125-preview', 1),  # GPT-4 Turbo
-    ('gpt-4-1106-preview', 1),  # GPT-4 Turbo
-    ('gpt-4-vision-preview', 1),  # GPT-4 Turbo with Vision.
-    # GPT-4 models.
-    ('gpt-4', 4),
-    ('gpt-4-0613', 4),
-    ('gpt-4-0314', 4),
-    ('gpt-4-32k', 4),
-    ('gpt-4-32k-0613', 4),
-    ('gpt-4-32k-0314', 4),
-    # GPT-3.5 Turbo models.
-    ('gpt-3.5-turbo', 16),
-    ('gpt-3.5-turbo-0125', 16),
-    ('gpt-3.5-turbo-1106', 16),
-    ('gpt-3.5-turbo-0613', 16),
-    ('gpt-3.5-turbo-0301', 16),
-    ('gpt-3.5-turbo-16k', 16),
-    ('gpt-3.5-turbo-16k-0613', 16),
-    ('gpt-3.5-turbo-16k-0301', 16),
-    # GPT-3.5 models.
-    ('text-davinci-003', 8),  # GPT-3.5, trained with RHLF.
-    ('text-davinci-002', 4),  # Trained with SFT but no RHLF.
-    ('code-davinci-002', 4),
-    # GPT-3 instruction-tuned models.
-    ('text-curie-001', 4),
-    ('text-babbage-001', 4),
-    ('text-ada-001', 4),
-    ('davinci', 4),
-    ('curie', 4),
-    ('babbage', 4),
-    ('ada', 4),
-    # GPT-3 base models without instruction tuning.
-    ('babbage-002', 4),
-    ('davinci-002', 4),
-]
-
-
-# Model concurreny setting.
-_MODEL_CONCURRENCY = {m[0]: m[1] for m in SUPPORTED_MODELS_AND_SETTINGS}
+# From https://platform.openai.com/settings/organization/limits
+_DEFAULT_TPM = 250000
+_DEFAULT_RPM = 3000
+
+SUPPORTED_MODELS_AND_SETTINGS = {
+    # Models from https://platform.openai.com/docs/models
+    # RPM is from https://platform.openai.com/docs/guides/rate-limits
+    # o1 (preview) models.
+    # Pricing in US dollars, from https://openai.com/api/pricing/
+    # as of 2024-10-10.
+    'o1': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=5000000,
+        cost_per_1k_input_tokens=0.015,
+        cost_per_1k_output_tokens=0.06,
+    ),
+    'o1-preview': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=5000000,
+        cost_per_1k_input_tokens=0.015,
+        cost_per_1k_output_tokens=0.06,
+    ),
+    'o1-preview-2024-09-12': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=5000000,
+        cost_per_1k_input_tokens=0.015,
+        cost_per_1k_output_tokens=0.06,
+    ),
+    'o1-mini': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=5000000,
+        cost_per_1k_input_tokens=0.003,
+        cost_per_1k_output_tokens=0.012,
+    ),
+    'o1-mini-2024-09-12': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=5000000,
+        cost_per_1k_input_tokens=0.003,
+        cost_per_1k_output_tokens=0.012,
+    ),
+    # GPT-4o models
+    'gpt-4o-mini': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=5000000,
+        cost_per_1k_input_tokens=0.00015,
+        cost_per_1k_output_tokens=0.0006,
+    ),
+    'gpt-4o-mini-2024-07-18': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=5000000,
+        cost_per_1k_input_tokens=0.00015,
+        cost_per_1k_output_tokens=0.0006,
+    ),
+    'gpt-4o': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=5000000,
+        cost_per_1k_input_tokens=0.0025,
+        cost_per_1k_output_tokens=0.01,
+    ),
+    'gpt-4o-2024-11-20': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=5000000,
+        cost_per_1k_input_tokens=0.0025,
+        cost_per_1k_output_tokens=0.01,
+    ),
+    'gpt-4o-2024-08-06': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=5000000,
+        cost_per_1k_input_tokens=0.0025,
+        cost_per_1k_output_tokens=0.01,
+    ),
+    'gpt-4o-2024-05-13': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=5000000,
+        cost_per_1k_input_tokens=0.005,
+        cost_per_1k_output_tokens=0.015,
+    ),
+    # GPT-4-Turbo models
+    'gpt-4-turbo': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.01,
+        cost_per_1k_output_tokens=0.03,
+    ),
+    'gpt-4-turbo-2024-04-09': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.01,
+        cost_per_1k_output_tokens=0.03,
+    ),
+    'gpt-4-turbo-preview': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.01,
+        cost_per_1k_output_tokens=0.03,
+    ),
+    'gpt-4-0125-preview': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.01,
+        cost_per_1k_output_tokens=0.03,
+    ),
+    'gpt-4-1106-preview': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.01,
+        cost_per_1k_output_tokens=0.03,
+    ),
+    'gpt-4-vision-preview': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.01,
+        cost_per_1k_output_tokens=0.03,
+    ),
+    'gpt-4-1106-vision-preview': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.01,
+        cost_per_1k_output_tokens=0.03,
+    ),
+    # GPT-4 models
+    'gpt-4': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=300000,
+        cost_per_1k_input_tokens=0.03,
+        cost_per_1k_output_tokens=0.06,
+    ),
+    'gpt-4-0613': pg.Dict(
+        in_service=False,
+        rpm=10000,
+        tpm=300000,
+        cost_per_1k_input_tokens=0.03,
+        cost_per_1k_output_tokens=0.06,
+    ),
+    'gpt-4-0314': pg.Dict(
+        in_service=False,
+        rpm=10000,
+        tpm=300000,
+        cost_per_1k_input_tokens=0.03,
+        cost_per_1k_output_tokens=0.06,
+    ),
+    'gpt-4-32k': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=300000,
+        cost_per_1k_input_tokens=0.06,
+        cost_per_1k_output_tokens=0.12,
+    ),
+    'gpt-4-32k-0613': pg.Dict(
+        in_service=False,
+        rpm=10000,
+        tpm=300000,
+        cost_per_1k_input_tokens=0.06,
+        cost_per_1k_output_tokens=0.12,
+    ),
+    'gpt-4-32k-0314': pg.Dict(
+        in_service=False,
+        rpm=10000,
+        tpm=300000,
+        cost_per_1k_input_tokens=0.06,
+        cost_per_1k_output_tokens=0.12,
+    ),
+    # GPT-3.5-Turbo models
+    'gpt-3.5-turbo': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.0005,
+        cost_per_1k_output_tokens=0.0015,
+    ),
+    'gpt-3.5-turbo-0125': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.0005,
+        cost_per_1k_output_tokens=0.0015,
+    ),
+    'gpt-3.5-turbo-1106': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.001,
+        cost_per_1k_output_tokens=0.002,
+    ),
+    'gpt-3.5-turbo-0613': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.0015,
+        cost_per_1k_output_tokens=0.002,
+    ),
+    'gpt-3.5-turbo-0301': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.0015,
+        cost_per_1k_output_tokens=0.002,
+    ),
+    'gpt-3.5-turbo-16k': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.003,
+        cost_per_1k_output_tokens=0.004,
+    ),
+    'gpt-3.5-turbo-16k-0613': pg.Dict(
+        in_service=True,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.003,
+        cost_per_1k_output_tokens=0.004,
+    ),
+    'gpt-3.5-turbo-16k-0301': pg.Dict(
+        in_service=False,
+        rpm=10000,
+        tpm=2000000,
+        cost_per_1k_input_tokens=0.003,
+        cost_per_1k_output_tokens=0.004,
+    ),
+    # GPT-3.5 models
+    'text-davinci-003': pg.Dict(
+        in_service=False, rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM
+    ),
+    'text-davinci-002': pg.Dict(
+        in_service=False, rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM
+    ),
+    'code-davinci-002': pg.Dict(
+        in_service=False, rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM
+    ),
+    # GPT-3 instruction-tuned models (Deprecated)
+    'text-curie-001': pg.Dict(
+        in_service=False, rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM
+    ),
+    'text-babbage-001': pg.Dict(
+        in_service=False,
+        rpm=_DEFAULT_RPM,
+        tpm=_DEFAULT_TPM,
+    ),
+    'text-ada-001': pg.Dict(
+        in_service=False,
+        rpm=_DEFAULT_RPM,
+        tpm=_DEFAULT_TPM,
+    ),
+    'davinci': pg.Dict(
+        in_service=False,
+        rpm=_DEFAULT_RPM,
+        tpm=_DEFAULT_TPM,
+    ),
+    'curie': pg.Dict(in_service=False, rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    'babbage': pg.Dict(in_service=False, rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    'ada': pg.Dict(in_service=False, rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    # GPT-3 base models that are still in service.
+    'babbage-002': pg.Dict(in_service=True, rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    'davinci-002': pg.Dict(in_service=True, rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+}
 
 
 @lf.use_init_args(['model'])
-class OpenAI(lf.LanguageModel):
+class OpenAI(openai_compatible.OpenAICompatible):
   """OpenAI model."""
 
   model: pg.typing.Annotated[
       pg.typing.Enum(
-          pg.MISSING_VALUE, [m[0] for m in SUPPORTED_MODELS_AND_SETTINGS]
+          pg.MISSING_VALUE, list(SUPPORTED_MODELS_AND_SETTINGS.keys())
       ),
       'The name of the model to use.',
-  ] = 'gpt-3.5-turbo'
+  ]
 
-  multimodal: Annotated[
-      bool,
-      'Whether this model has multimodal support.'
-  ] = False
+  api_endpoint: str = 'https://api.openai.com/v1/chat/completions'
 
   api_key: Annotated[
       str | None,
@@ -120,23 +327,44 @@ class OpenAI(lf.LanguageModel):
       ),
   ] = None
 
+  project: Annotated[
+      str | None,
+      (
+          'Project. If None, the key will be read from environment '
+          "variable 'OPENAI_PROJECT'. Based on the value, usages from "
+          "these API requests will count against the project's quota. "
+      ),
+  ] = None
+
   def _on_bound(self):
     super()._on_bound()
-    self.__dict__.pop('_api_initialized', None)
+    self._api_key = None
+    self._organization = None
+    self._project = None
 
-  @functools.cached_property
-  def _api_initialized(self):
+  def _initialize(self):
     api_key = self.api_key or os.environ.get('OPENAI_API_KEY', None)
     if not api_key:
       raise ValueError(
           'Please specify `api_key` during `__init__` or set environment '
           'variable `OPENAI_API_KEY` with your OpenAI API key.'
       )
-    openai.api_key = api_key
-    org = self.organization or os.environ.get('OPENAI_ORGANIZATION', None)
-    if org:
-      openai.organization = org
-    return True
+    self._api_key = api_key
+    self._organization = self.organization or os.environ.get(
+        'OPENAI_ORGANIZATION', None
+    )
+    self._project = self.project or os.environ.get('OPENAI_PROJECT', None)
+
+  @property
+  def headers(self) -> dict[str, Any]:
+    assert self._api_initialized
+    headers = super().headers
+    headers['Authorization'] = f'Bearer {self._api_key}'
+    if self._organization:
+      headers['OpenAI-Organization'] = self._organization
+    if self._project:
+      headers['OpenAI-Project'] = self._project
+    return headers
 
   @property
   def model_id(self) -> str:
@@ -145,144 +373,73 @@ class OpenAI(lf.LanguageModel):
 
   @property
   def max_concurrency(self) -> int:
-    return _MODEL_CONCURRENCY[self.model]
+    rpm = SUPPORTED_MODELS_AND_SETTINGS[self.model].get('rpm', 0)
+    tpm = SUPPORTED_MODELS_AND_SETTINGS[self.model].get('tpm', 0)
+    return self.rate_to_max_concurrency(
+        requests_per_min=rpm, tokens_per_min=tpm
+    )
+
+  def estimate_cost(
+      self,
+      num_input_tokens: int,
+      num_output_tokens: int
+  ) -> float | None:
+    """Estimate the cost based on usage."""
+    cost_per_1k_input_tokens = SUPPORTED_MODELS_AND_SETTINGS[self.model].get(
+        'cost_per_1k_input_tokens', None
+    )
+    cost_per_1k_output_tokens = SUPPORTED_MODELS_AND_SETTINGS[self.model].get(
+        'cost_per_1k_output_tokens', None
+    )
+    if cost_per_1k_output_tokens is None or cost_per_1k_input_tokens is None:
+      return None
+    return (
+        cost_per_1k_input_tokens * num_input_tokens
+        + cost_per_1k_output_tokens * num_output_tokens
+    ) / 1000
 
   @classmethod
   def dir(cls):
-    return openai.Model.list()
-
-  @property
-  def is_chat_model(self):
-    """Returns True if the model is a chat model."""
-    return self.model.startswith(('gpt-4', 'gpt-3.5-turbo'))
+    return [k for k, v in SUPPORTED_MODELS_AND_SETTINGS.items() if v.in_service]
 
-  def _get_request_args(
+  def _request_args(
       self, options: lf.LMSamplingOptions) -> dict[str, Any]:
-    # Reference:
-    # https://platform.openai.com/docs/api-reference/completions/create
-    # NOTE(daiyip): options.top_k is not applicable.
-    args = dict(
-        n=options.n,
-        temperature=options.temperature,
-        max_tokens=options.max_tokens,
-        stream=False,
-        timeout=self.timeout,
-        logprobs=options.logprobs,
-        top_logprobs=options.top_logprobs,
-    )
-    # Completion and ChatCompletion uses different parameter name for model.
-    args['model' if self.is_chat_model else 'engine'] = self.model
+    # Reasoning models (o1 series) does not support `logprobs` by 2024/09/12.
+    if options.logprobs and self.model.startswith(('o1-', 'o3-')):
+      raise RuntimeError('`logprobs` is not supported on {self.model!r}.')
+    return super()._request_args(options)
 
-    if options.top_p is not None:
-      args['top_p'] = options.top_p
-    if options.stop:
-      args['stop'] = options.stop
-    return args
 
-  def _sample(self, prompts: list[lf.Message]) -> list[LMSamplingResult]:
-    assert self._api_initialized
-    if self.is_chat_model:
-      return self._chat_complete_batch(prompts)
-    else:
-      return self._complete_batch(prompts)
-
-  def _complete_batch(
-      self, prompts: list[lf.Message]) -> list[LMSamplingResult]:
-
-    def _open_ai_completion(prompts):
-      response = openai.Completion.create(
-          prompt=[p.text for p in prompts],
-          **self._get_request_args(self.sampling_options),
-      )
-      response = cast(openai_object.OpenAIObject, response)
-      # Parse response.
-      samples_by_index = collections.defaultdict(list)
-      for choice in response.choices:
-        samples_by_index[choice.index].append(
-            lf.LMSample(choice.text.strip(), score=choice.logprobs or 0.0)
-        )
-
-      usage = Usage(
-          prompt_tokens=response.usage.prompt_tokens,
-          completion_tokens=response.usage.completion_tokens,
-          total_tokens=response.usage.total_tokens,
-      )
-      return [
-          LMSamplingResult(
-              samples_by_index[index], usage=usage if index == 0 else None
-          )
-          for index in sorted(samples_by_index.keys())
-      ]
-
-    return self._parallel_execute_with_currency_control(
-        _open_ai_completion,
-        [prompts],
-        retry_on_errors=(
-            openai_error.ServiceUnavailableError,
-            openai_error.RateLimitError,
-        ),
-    )[0]
-
-  def _chat_complete_batch(
-      self, prompts: list[lf.Message]
-  ) -> list[LMSamplingResult]:
-    def _open_ai_chat_completion(prompt: lf.Message):
-      if self.multimodal:
-        content = []
-        for chunk in prompt.chunk():
-          if isinstance(chunk, str):
-            item = dict(type='text', text=chunk)
-          elif isinstance(chunk, lf_modalities.Image) and chunk.uri:
-            item = dict(type='image_url', image_url=chunk.uri)
-          else:
-            raise ValueError(f'Unsupported modality object: {chunk!r}.')
-          content.append(item)
-      else:
-        content = prompt.text
-
-      response = openai.ChatCompletion.create(
-          # TODO(daiyip): support conversation history and system prompt.
-          messages=[{'role': 'user', 'content': content}],
-          **self._get_request_args(self.sampling_options),
-      )
-      response = cast(openai_object.OpenAIObject, response)
-      samples = []
-      for choice in response.choices:
-        logprobs = None
-        if choice.logprobs:
-          logprobs = [
-              (
-                  t.token,
-                  t.logprob,
-                  [(tt.token, tt.logprob) for tt in t.top_logprobs],
-              )
-              for t in choice.logprobs.content
-          ]
-        samples.append(
-            lf.LMSample(
-                choice.message.content,
-                score=0.0,
-                logprobs=logprobs,
-            )
-        )
-
-      return LMSamplingResult(
-          samples=samples,
-          usage=Usage(
-              prompt_tokens=response.usage.prompt_tokens,
-              completion_tokens=response.usage.completion_tokens,
-              total_tokens=response.usage.total_tokens,
-          ),
-      )
+class GptO1(OpenAI):
+  """GPT-O1."""
 
-    return self._parallel_execute_with_currency_control(
-        _open_ai_chat_completion,
-        prompts,
-        retry_on_errors=(
-            openai_error.ServiceUnavailableError,
-            openai_error.RateLimitError,
-        ),
-    )
+  model = 'o1'
+  multimodal = True
+  timeout = None
+
+
+class GptO1Preview(OpenAI):
+  """GPT-O1."""
+  model = 'o1-preview'
+  timeout = None
+
+
+class GptO1Preview_20240912(OpenAI):   # pylint: disable=invalid-name
+  """GPT O1."""
+  model = 'o1-preview-2024-09-12'
+  timeout = None
+
+
+class GptO1Mini(OpenAI):
+  """GPT O1-mini."""
+  model = 'o1-mini'
+  timeout = None
+
+
+class GptO1Mini_20240912(OpenAI):   # pylint: disable=invalid-name
+  """GPT O1-mini."""
+  model = 'o1-mini-2024-09-12'
+  timeout = None
 
 
 class Gpt4(OpenAI):
@@ -291,27 +448,44 @@ class Gpt4(OpenAI):
 
 
 class Gpt4Turbo(Gpt4):
-  """GPT-4 Turbo with 128K context window size. Knowledge up to 4-2023."""
-  model = 'gpt-4-turbo-preview'
+  """GPT-4 Turbo with 128K context window. Knowledge up to Dec. 2023."""
+  model = 'gpt-4-turbo'
+  multimodal = True
 
 
-class Gpt4TurboVision(Gpt4Turbo):
-  """GPT-4 Turbo with vision."""
-  model = 'gpt-4-vision-preview'
+class Gpt4Turbo_20240409(Gpt4Turbo):  # pylint:disable=invalid-name
+  """GPT-4 Turbo with 128K context window. Knowledge up to Dec. 2023."""
+  model = 'gpt-4-turbo-2024-04-09'
   multimodal = True
 
 
-class Gpt4Turbo_0125(Gpt4Turbo):   # pylint:disable=invalid-name
-  """GPT-4 Turbo with 128K context window size. Knowledge up to 4-2023."""
+class Gpt4TurboPreview(Gpt4):
+  """GPT-4 Turbo Preview with 128k context window. Knowledge up to Dec. 2023."""
+  model = 'gpt-4-turbo-preview'
+
+
+class Gpt4TurboPreview_20240125(Gpt4TurboPreview):  # pylint: disable=invalid-name
+  """GPT-4 Turbo Preview with 128k context window. Knowledge up to Dec. 2023."""
   model = 'gpt-4-0125-preview'
 
 
-class Gpt4Turbo_1106(Gpt4Turbo):   # pylint:disable=invalid-name
-  """GPT-4 Turbo @20231106. 128K context window. Knowledge up to 4-2023."""
+class Gpt4TurboPreview_20231106(Gpt4TurboPreview):  # pylint: disable=invalid-name
+  """GPT-4 Turbo Preview with 128k context window. Knowledge up to Apr. 2023."""
   model = 'gpt-4-1106-preview'
 
 
-class Gpt4_0613(Gpt4):    # pylint:disable=invalid-name
+class Gpt4VisionPreview(Gpt4):
+  """GPT-4 Turbo vision preview. 128k context window. Knowledge to Apr. 2023."""
+  model = 'gpt-4-vision-preview'
+  multimodal = True
+
+
+class Gpt4VisionPreview_20231106(Gpt4):  # pylint: disable=invalid-name
+  """GPT-4 Turbo vision preview. 128k context window. Knowledge to Apr. 2023."""
+  model = 'gpt-4-1106-vision-preview'
+
+
+class Gpt4_20230613(Gpt4):    # pylint:disable=invalid-name
   """GPT-4 @20230613. 8K context window. Knowledge up to 9-2021."""
   model = 'gpt-4-0613'
 
@@ -321,11 +495,47 @@ class Gpt4_32K(Gpt4):       # pylint:disable=invalid-name
   model = 'gpt-4-32k'
 
 
-class Gpt4_32K_0613(Gpt4_32K):    # pylint:disable=invalid-name
+class Gpt4_32K_20230613(Gpt4_32K):    # pylint:disable=invalid-name
   """GPT-4 @20230613. 32K context window. Knowledge up to 9-2021."""
   model = 'gpt-4-32k-0613'
 
 
+class Gpt4oMini(OpenAI):
+  """GPT-4o Mini."""
+  model = 'gpt-4o-mini'
+  multimodal = True
+
+
+class Gpt4oMini_20240718(OpenAI):  # pylint:disable=invalid-name
+  """GPT-4o Mini."""
+  model = 'gpt-4o-mini-2024-07-18'
+  multimodal = True
+
+
+class Gpt4o(OpenAI):
+  """GPT-4o."""
+  model = 'gpt-4o'
+  multimodal = True
+
+
+class Gpt4o_20241120(OpenAI):     # pylint:disable=invalid-name
+  """GPT-4o version 2024-11-20."""
+  model = 'gpt-4o-2024-11-20'
+  multimodal = True
+
+
+class Gpt4o_20240806(OpenAI):     # pylint:disable=invalid-name
+  """GPT-4o version 2024-08-06."""
+  model = 'gpt-4o-2024-08-06'
+  multimodal = True
+
+
+class Gpt4o_20240513(OpenAI):     # pylint:disable=invalid-name
+  """GPT-4o version 2024-05-13."""
+  model = 'gpt-4o-2024-05-13'
+  multimodal = True
+
+
 class Gpt35(OpenAI):
   """GPT-3.5. 4K max tokens, trained up on data up to Sep, 2021."""
   model = 'text-davinci-003'
@@ -336,17 +546,17 @@ class Gpt35Turbo(Gpt35):
   model = 'gpt-3.5-turbo'
 
 
-class Gpt35Turbo_0125(Gpt35Turbo):   # pylint:disable=invalid-name
+class Gpt35Turbo_20240125(Gpt35Turbo):   # pylint:disable=invalid-name
   """GPT-3.5 Turbo @20240125. 16K context window. Knowledge up to 09/2021."""
   model = 'gpt-3.5-turbo-0125'
 
 
-class Gpt35Turbo_1106(Gpt35Turbo):   # pylint:disable=invalid-name
+class Gpt35Turbo_20231106(Gpt35Turbo):   # pylint:disable=invalid-name
   """Gpt3.5 Turbo @20231106. 16K context window. Knowledge up to 09/2021."""
   model = 'gpt-3.5-turbo-1106'
 
 
-class Gpt35Turbo_0613(Gpt35Turbo):   # pylint:disable=invalid-name
+class Gpt35Turbo_20230613(Gpt35Turbo):   # pylint:disable=invalid-name
   """Gpt3.5 Turbo snapshot at 2023/06/13, with 4K context window size."""
   model = 'gpt-3.5-turbo-0613'
 
@@ -356,7 +566,7 @@ class Gpt35Turbo16K(Gpt35Turbo):
   model = 'gpt-3.5-turbo-16k'
 
 
-class Gpt35Turbo16K_0613(Gpt35Turbo):   # pylint:disable=invalid-name
+class Gpt35Turbo16K_20230613(Gpt35Turbo):   # pylint:disable=invalid-name
   """Gtp 3.5 Turbo 16K 0613."""
   model = 'gpt-3.5-turbo-16k-0613'