litellm - 💡(How to fix) Fix [Bug]: SambaNova cost map is stale/incorrect (deprecated models, wrong pricing/context, missing models)

Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

What happened?

Of the 17 sambanova/* entries in model_prices_and_context_window.json, 12 are already retired on SambaNova Cloud but remain in the cost map with no deprecation_date. Several current models also carry incorrect pricing/context, and 2 current models are missing.

Official references: models / deprecations / pricing

Problem 1: deprecated models still present (12, none has deprecation_date)

Problem 2: current models with inaccurate metadata (prices per 1M tokens, from the official pricing page)

(gpt-oss-120b currently carries DeepSeek-V3.1's pricing, likely a copy-paste when both were added in #14866.)

Problem 3: current models missing (2)

Expected: the sambanova/* entries reflect SambaNova Cloud: deprecated models tagged with deprecation_date (matching the existing convention, 58 entries already use it), current models have accurate pricing/context, and the missing current models are added.

Steps to Reproduce

import litellm

# Problem 2: pricing / context disagree with SambaNova's official pricing page
print(litellm.model_cost["sambanova/gpt-oss-120b"]["input_cost_per_token"])    # 3e-06  (official 2.2e-07)
print(litellm.model_cost["sambanova/gpt-oss-120b"]["output_cost_per_token"])   # 4.5e-06 (official 5.9e-07)
print(litellm.model_cost["sambanova/MiniMax-M2.7"]["input_cost_per_token"])    # 3e-07  (official 6e-07)
print(litellm.model_cost["sambanova/DeepSeek-V3.1"]["max_input_tokens"])       # 32768  (official 131072)

# Problem 1: retired models still present, none tagged with deprecation_date
sn = {k: v for k, v in litellm.model_cost.items() if v.get("litellm_provider") == "sambanova"}
print(len(sn))                                                                 # 17
print("sambanova/Meta-Llama-3.1-405B-Instruct" in sn)                          # True (retired 2025-06-25)
print(any(v.get("deprecation_date") for v in sn.values()))                     # False

# Problem 3: current models missing
print("sambanova/DeepSeek-V3.2" in sn, "sambanova/gemma-3-12b-it" in sn)       # False False

Relevant log output

# (optional) output of the snippet above

What part of LiteLLM is this about?

SDK (litellm Python package)

What LiteLLM version are you on?

v1.83.14-stable (also reproduces on main)

Proposed fix

• Update both model_prices_and_context_window.json and litellm/model_prices_and_context_window_backup.json identically (ci_cd/check_files_match.py enforces a byte match).
• Add deprecation_date to the 12 deprecated entries (do not delete them; same convention as the 58 entries already tagged).
• Fix pricing/context for gpt-oss-120b, MiniMax-M2.7, DeepSeek-V3.1.
• Add DeepSeek-V3.2 and gemma-3-12b-it.

Notes:

• #20497 partially overlaps but adds no deprecation_date, does not update the backup JSON, and adds models that are themselves already deprecated (MiniMax-M2.5, DeepSeek-V3.1-Terminus, ALLaM-7B); out of scope here.
• The official pricing page also lists DeepSeek-V3.1-cb ($0.15 / $0.75), but it is not on the SambaCloud models page, so it is left out (needs confirmation).
• The fact that these retired models still appear in the Admin UI "Add Model" dropdown is filed separately as a UI Dashboard issue (#29012; the dropdown does not filter on deprecation_date).