llm-eol

Scrapes LLM model deprecation and retirement dates from provider documentation, matches them against your list of active models, maintains a persistent local database, and exports results to Google Sheets with risk-based colour coding.

Before You Start (please read)

To run this on your own computer you need one file: credentials.json. This is the key that lets the script sign in to Google and write to our shared sheet. For security, everyone uses their own key — it is never passed around between people, and it is never committed to this project (git ignores it).

Two one-time steps:

1. Create your own credentials.json. Follow Setup steps 3–4 below: in a Google Cloud project turn on the Google Sheets and Drive APIs, create a service account, and download its JSON key. Save that file in the main folder of this project (the same folder as this README) and name it exactly credentials.json. The script finds it automatically.

2. Get your key access to the shared sheet. Open your credentials.json and copy the client_email value — it looks like something@your-project.iam.gserviceaccount.com. Send that address to Ben (bnguyen@studiosity.com), and he'll share the team sheet with it so your key is allowed to write. (You only do this once.)

We all write to the same Google Sheet — on purpose. The sheet is already set in the code, so everyone's results go to one shared place. Please do not change the sheet ID in src/main.py and do not create your own sheet — if we each used a different one we'd end up with scattered, out-of-date data instead of one view everyone can trust. Once your key has access, run the script whenever you like to keep the shared sheet current (see Usage).

Supported Providers

Provider	Source page
Google Gemini	ai.google.dev/gemini-api/docs/deprecations
OpenAI	developers.openai.com/api/docs/deprecations
Azure OpenAI	learn.microsoft.com — model retirements
Anthropic	platform.claude.com/docs/about-claude/model-deprecations
Vertex AI	docs.cloud.google.com — partner models deprecations
AWS Bedrock	docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html

AWS Bedrock model card pages (context window, modalities, geo inference IDs, etc.) are also scraped from the provider-specific card hierarchy under model-cards.html.

How It Works

1. Scrapes each provider's deprecation/lifecycle page
2. Scrapes AWS Bedrock individual model card pages for rich metadata
3. Merges results into data/models_db.json — records are never deleted by a scrape run (models that disappear from a provider page keep their last-known data)
4. Prunes records whose shutdown date expired more than 1 year ago
5. Matches the full DB against MY_MODELS in src/main.py
6. Exports three Google Sheets tabs with colour-coded risk levels

Risk Levels

Level	Condition	Colour
EXPIRED	Already past shutdown date	Muted rose
CRITICAL	≤ 30 days remaining	Soft peach-orange
HIGH	≤ 90 days remaining	Soft amber
MEDIUM	≤ 180 days remaining	Pale yellow
LOW	> 180 days remaining	Soft mint
Unknown	Date could not be parsed	White
Not found	Model not in any provider page	Light grey

Google Sheets Output

All Models (tab 1)

Every record in the local DB, across all providers.

Column	Notes
Provider
Model	Model ID or human-readable name
Lifecycle Stage	Active / Legacy / EOL — AWS Bedrock only
Scraped Shutdown Date	Raw string from the provider page
Parsed Shutdown Date	Normalised to YYYY-MM-DD; N/A if unparseable
Days Remaining	Integer; negative = already expired
Risk Level	See table above; colour-coded cell
Source URL	Direct link to the provider page
First Seen	Date the record was first added to the local DB
Last Seen	Date the record was last confirmed by a scrape run

Interested Models (tab 2)

Only models from MY_MODELS. Unmatched models appear at the bottom in grey with "Not found" values so nothing is silently omitted.

Column	Notes
Last Updated	Timestamp of the run (Australia/Melbourne timezone)
Our Model	Exactly as written in MY_MODELS
Scraped Model	Matched identifier from the provider page
Provider
Scraped Shutdown Date
Parsed Shutdown Date
Days Remaining
Risk Level	Colour-coded cell

Bedrock Details (tab 3)

AWS Bedrock models that have model card metadata. Only shown when card data is available.

Column	Notes
Model ID
Lifecycle Stage
Context Window	Tokens (integer)
Max Output Tokens	Tokens (integer)
Input Modalities	Comma-separated list
Output Modalities	Comma-separated list
Knowledge Cutoff
Geo Inference IDs	Comma-separated cross-region inference profile IDs
Model Card URL	Direct link to the AWS Bedrock model card page

Configuration

Edit src/main.py:

# Target Google Sheet
SPREADSHEET_ID = 'your-sheet-id-from-url'

# Models your application uses
MY_MODELS = [
    "gpt-4o-mini",
    "anthropic.claude-3-haiku-20240307-v1:0",
    "us.meta.llama3-3-70b-instruct-v1:0",   # Bedrock cross-region prefix stripped automatically
    ...
]

MY_MODELS matching rules (applied in order):

1. Exact match
2. Scraped model has appended version info — gpt-4o matches gpt-4o (2024-05-13)
3. User model has appended version tag — claude-3-haiku@20240307 matches claude-3-haiku
4. AWS Bedrock cross-region prefix stripped — us.meta.llama3-... matches meta.llama3-... Supported prefixes: us. eu. ap. apac. au. ca. jp. global. us-gov.

Usage

python src/main.py

Run from the project root. The src/ directory is automatically on the Python path.

Setup

1. Python Version

Python 3.8 or later is required.

2. Install Dependencies

pip install -r requirements.txt

3. Enable Google Sheets API

1. Go to Google Cloud Console
2. Create or select a project
3. Enable the Google Sheets API and Google Drive API

4. Create a Service Account

1. Navigate to IAM & Admin > Service Accounts
2. Click Create Service Account (e.g. llm-eol-tracker)
3. Click the account → Keys tab → Add Key > Create New Key > JSON
4. Save the downloaded file as credentials.json in the project root, or set GOOGLE_CREDENTIALS_FILE to its path

credentials.json is listed in .gitignore and must never be committed.

5. Share the Spreadsheet

Copy the service account email from credentials.json and share your target Google Sheet with it (Editor access).

Local Database

Model records are persisted to data/models_db.json. This file:

• Is safe to commit — it contains no secrets, only scraped public data
• Grows incrementally — records are merged in on each run, never bulk-replaced
• Retains records that disappear from provider pages, with last_seen showing when they were last confirmed
• Has expired entries (shutdown date > 1 year ago) pruned automatically on each run

The Google Sheet is a human-readable view of the database, not a replacement for it. It is output-only; the JSON file is the operational source of truth.

Project Layout

src/
  main.py                    ← entry point: MY_MODELS, SPREADSHEET_ID, run order
  utils.py                   ← get_html, parse_shutdown_date, calculate_risk_info
  checker.py                 ← check_my_models, Bedrock geo-prefix matching
  sheets.py                  ← Google Sheets export (3 tabs)
  database.py                ← load/save/merge/cleanup for data/models_db.json
  parsers/
    __init__.py              ← parse_all_deprecations (calls all parsers, deduplicates)
    google_gemini.py
    openai.py
    azure_openai.py
    anthropic.py
    vertex_ai.py
    bedrock.py               ← lifecycle page (Active / Legacy / EOL tables)
    bedrock_model_cards.py   ← model card pages (context window, modalities, etc.)
data/
  models_db.json             ← persistent model database (auto-created on first run)