"""PTU model data table for Azure OpenAI provisioned throughput sizing.
All values are sourced from the official Microsoft documentation:
https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/provisioned-throughput-onboarding
Data should be updated when the upstream doc table changes.
"""
# Version of this embedded data table — bump when updating from docs.
DATA_VERSION = "2026-02-01"
# Authoritative source URL for PTU model tables.
DATA_SOURCE_URL = (
"https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/" "provisioned-throughput-onboarding"
)
# Max PTUs per single deployment (hard limit from Azure quotas/limits docs).
MAX_PTUS_PER_DEPLOYMENT = 100_000
# ---------------------------------------------------------------------------
# Deployment type metadata
# ---------------------------------------------------------------------------
DEPLOYMENT_TYPES: dict[str, dict[str, str]] = {
"GlobalProvisioned": {
"label": "Global Provisioned",
"description": "Traffic may be processed in any Azure geography. Best for highest capacity.",
"min_key": "global_min_ptus",
"increment_key": "global_increment",
},
"DataZoneProvisioned": {
"label": "Data Zone Provisioned",
"description": "Traffic processed within a defined data zone (e.g., EU, US). Shares PTU table with Global.",
"min_key": "global_min_ptus",
"increment_key": "global_increment",
},
"RegionalProvisioned": {
"label": "Regional Provisioned",
"description": "Traffic processed in a single Azure region. Stricter data residency.",
"min_key": "regional_min_ptus",
"increment_key": "regional_increment",
},
}
# ---------------------------------------------------------------------------
# Per-model PTU characteristics
# ---------------------------------------------------------------------------
# Each entry must contain:
# input_tpm_per_ptu – Input tokens-per-minute capacity per PTU
# output_multiplier – 1 output token counts as N input tokens for utilization
# global_min_ptus – Minimum PTUs for Global / Data Zone deployments
# global_increment – Scale increment for Global / Data Zone deployments
# regional_min_ptus – Minimum PTUs for Regional deployments (None if unsupported)
# regional_increment – Scale increment for Regional deployments (None if unsupported)
#
# output_multiplier notes:
# Docs state: "Starting with GPT 4.1 models and later, the system generally
# matches the global standard price ratio between input and output tokens."
# - gpt-5 family: explicitly documented as 8× (1 output = 8 input tokens)
# - gpt-4.1 family: explicitly documented as 4× (1 output = 4 input tokens)
# - Llama-3.3-70B-Instruct: explicitly documented as 4× (exception to pricing ratio)
# - Previous Azure OpenAI models (gpt-4o, gpt-4o-mini): 3× (verified via
# Foundry calculator and official MS docs example tables)
# - o3-mini, o1: assumed 3× (same "previous model" category; docs say
# "older models use a different ratio" without specifying)
# ---------------------------------------------------------------------------
PTU_MODEL_TABLE: dict[str, dict] = {
# ── Latest Azure OpenAI models ──────────────────────────────────────
"gpt-5.2": {
"input_tpm_per_ptu": 3_400,
"output_multiplier": 8,
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 50,
"regional_increment": 50,
},
"gpt-5.2-codex": {
"input_tpm_per_ptu": 3_400,
"output_multiplier": 8,
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 50,
"regional_increment": 50,
},
"gpt-5.1": {
"input_tpm_per_ptu": 4_750,
"output_multiplier": 8,
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 50,
"regional_increment": 50,
},
"gpt-5.1-codex": {
"input_tpm_per_ptu": 4_750,
"output_multiplier": 8,
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 50,
"regional_increment": 50,
},
"gpt-5": {
"input_tpm_per_ptu": 4_750,
"output_multiplier": 8, # Docs: "for gpt-5, 1 output = 8 input tokens"
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 50,
"regional_increment": 50,
},
"gpt-5-mini": {
"input_tpm_per_ptu": 23_750,
"output_multiplier": 4,
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 25,
"regional_increment": 25,
},
"gpt-4.1": {
"input_tpm_per_ptu": 3_000,
"output_multiplier": 4, # Docs: "for gpt-4.1, 1 output = 4 input tokens"
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 50,
"regional_increment": 50,
},
"gpt-4.1-mini": {
"input_tpm_per_ptu": 14_900,
"output_multiplier": 4,
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 25,
"regional_increment": 25,
},
"gpt-4.1-nano": {
"input_tpm_per_ptu": 59_400,
"output_multiplier": 4,
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 25,
"regional_increment": 25,
},
"o3": {
"input_tpm_per_ptu": 3_000,
"output_multiplier": 4,
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 50,
"regional_increment": 50,
},
"o4-mini": {
"input_tpm_per_ptu": 5_400,
"output_multiplier": 4,
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 25,
"regional_increment": 25,
},
# ── Previous Azure OpenAI models ────────────────────────────────────
"gpt-4o": {
"input_tpm_per_ptu": 2_500,
"output_multiplier": 3, # Verified via Foundry calculator; older model, different ratio
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 50,
"regional_increment": 50,
},
"gpt-4o-mini": {
"input_tpm_per_ptu": 37_000,
"output_multiplier": 3, # Verified via official MS docs example table (latency page)
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 25,
"regional_increment": 25,
},
"o3-mini": {
"input_tpm_per_ptu": 2_500,
"output_multiplier": 3, # Previous model; docs: "older models use a different ratio"
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 25,
"regional_increment": 25,
},
"o1": {
"input_tpm_per_ptu": 230,
"output_multiplier": 3, # Previous model; docs: "older models use a different ratio"
"global_min_ptus": 15,
"global_increment": 5,
"regional_min_ptus": 25,
"regional_increment": 50,
},
# ── Direct from Azure models ────────────────────────────────────────
"Llama-3.3-70B-Instruct": {
"input_tpm_per_ptu": 8_450,
"output_multiplier": 4, # Docs: explicit exception to pricing ratio
"global_min_ptus": 100,
"global_increment": 100,
"regional_min_ptus": None, # Regional not available
"regional_increment": None,
},
"DeepSeek-R1": {
"input_tpm_per_ptu": 4_000,
"output_multiplier": 4,
"global_min_ptus": 100,
"global_increment": 100,
"regional_min_ptus": None,
"regional_increment": None,
},
"DeepSeek-V3-0324": {
"input_tpm_per_ptu": 4_000,
"output_multiplier": 4,
"global_min_ptus": 100,
"global_increment": 100,
"regional_min_ptus": None,
"regional_increment": None,
},
"DeepSeek-R1-0528": {
"input_tpm_per_ptu": 4_000,
"output_multiplier": 4,
"global_min_ptus": 100,
"global_increment": 100,
"regional_min_ptus": None,
"regional_increment": None,
},
}
def get_supported_models() -> list[str]:
"""Return sorted list of supported model IDs."""
return sorted(PTU_MODEL_TABLE.keys())
def get_model_info(model: str) -> dict | None:
"""Look up a model in the PTU table (case-insensitive).
Returns the model entry dict or None if not found.
"""
# Exact match first
if model in PTU_MODEL_TABLE:
return PTU_MODEL_TABLE[model]
# Case-insensitive fallback
lower = model.lower()
for key, value in PTU_MODEL_TABLE.items():
if key.lower() == lower:
return value
return None
def get_model_canonical_name(model: str) -> str | None:
"""Return the canonical (case-correct) model name, or None if not found."""
if model in PTU_MODEL_TABLE:
return model
lower = model.lower()
for key in PTU_MODEL_TABLE:
if key.lower() == lower:
return key
return None
