How to Track OpenAI Updates: API Changelog, ChatGPT Release Notes, and Official Announcements Done Right

Tracking OpenAI updates is the ongoing practice of systematically reviewing OpenAI's three official surfaces and translating model releases, price changes, and retirements into engineering actions. OpenAI has never published updates from a single place: models and pricing move through the API changelog, product features and product-side retirements move through ChatGPT Release Notes, and company-level events move through official announcements -- miss any one surface and you can miss a retirement countdown that has already started. This guide maps the division of labor, explains the 90-day ChatGPT retirement convention versus API deprecations, and provides a weekly review SOP you can actually run.

1. Why You Need a Tracking Mechanism

Two reasons make "occasionally scrolling social media" inadequate. First, updates are scattered: model releases in the changelog, feature launches in the release notes, supply-relationship changes in announcements -- three places, different cadences, different wording. Second, retirements are accelerating: ChatGPT-side models now follow a "roughly 90 days after the successor ships" convention -- in 2026 alone, GPT-5.2 (6/12), GPT-4.5 (6/26), and o3 (8/26) left ChatGPT, with the shortest window from announcement to shutdown at 30 days. For teams with a hardcoded model name in production configs, this is not a news problem -- it is an incident waiting to happen.

2. The Division of Labor Across Three Surfaces

Each surface calls for a different reader action: the changelog maps directly to engineering changes (model names, parameters, price tables); Release Notes maps to user communication and training; company announcements map to vendor-risk assessment. A textbook example is DALL·E's two-layer shutdown -- the API models were hard-removed on 2026-05-12 (changelog), while the official DALL·E GPT inside ChatGPT only retires on 2026-08-30 (Release Notes): anyone watching a single surface gets blindsided on the other.

3. The ChatGPT Side: The 90-Day Retirement Convention

The official notes confirmed the convention in an entry dated 2026-06-12: after a successor is released, models generally remain available in ChatGPT for about 90 days. The 2026 track record:

Three practical corollaries: first, the countdown starts on the successor's launch date -- tracking "new model releases" is tracking "old model retirements"; second, the window can be as short as 30 days, so migration plans must be executable within a month; third, retirement announcements usually include a migration path (the DALL·E GPT notice points to ChatGPT Images) -- cite the official path in your user communications.

4. The API Side: Deprecation and Hard Removal Are Different Things

API removals follow the deprecation entries in the API changelog, on a cadence entirely separate from the ChatGPT product side. Two key distinctions:

• Product retirement does not mean API removal: when GPT-4o, GPT-4.1, GPT-4.1 mini, o4-mini, and GPT-5 retired from ChatGPT on 2026-02-13, the official notes stated that API availability remained unchanged. Production systems do not need an emergency migration because of a ChatGPT-side retirement -- but that is not a forever guarantee either; the deprecation notice is what counts.
• Hard removals have no buffer: DALL·E 2/3 were directly removed from the API on 2026-05-12, without a deprecation grace period. For entries like that, migration must complete before the effective date -- "change the model name on the day" is the only safe rhythm.

Team model inventories should therefore carry two fields: chatgpt_retirement (affects product communication) and api_deprecation (affects production code). Merging them guarantees errors.

5. Team Rollout: The Weekly Review SOP

Script the review around "snapshot + grep + diff." The following commands are tested against the live changelog (the page returns full HTML to curl, roughly 480KB):

Step 1: take a dated snapshot (put it in a weekly cron)

curl -s https://developers.openai.com/api/docs/changelog -o "changelog-$(date +%F).html"

Step 2: grep for retirement keywords and dates

grep -coE 'deprecat|retire|removed' changelog-2026-08-29.html
grep -oE '20[0-9]{2}-[0-9]{2}-[0-9]{2}' changelog-2026-08-29.html | sort -u | head

Step 3: diff against last week's snapshot for increments

diff changelog-2026-08-22.html changelog-2026-08-29.html | head -40

Two caveats: part of the changelog's recent content renders client-side, so what curl gets is a static skeleton -- grepping is for detecting "did anything change," and full reading still belongs in a browser. And the help.openai.com page hosting ChatGPT Release Notes has anti-scraping protection -- review it manually rather than fighting it with scripts.

6. Common Mistakes

Treating product retirement as API removal. Panicking and rewriting production configs because "model X left ChatGPT" -- most of the time the API is unchanged. The reverse also holds: no deprecation entry in the changelog does not mean there is no countdown on the product side.

Relying only on secondhand sources. Social media and digest newsletters are fast but inaccurate -- paraphrasing drops conditions and shifts dates. Use secondhand reports as leads and the three official surfaces as evidence.

Watching docs but not bills. The authoritative sources for price changes are the changelog and pricing pages, but your actual invoice is the final truth: cached pricing, tiered pricing, and batch discounts can all pull real costs away from list prices. When the review surfaces a price change, reconcile against the bill first.

Reviews that produce no actions. Read, logged, forgotten -- equivalent to no review. Every increment must land in one of three actions: change a config, schedule a task, or update a budget.

Frequently Asked Questions

1. Where should I watch for OpenAI updates?

Three official surfaces: the API changelog (developers.openai.com/api/docs/changelog) for model releases, pricing, and deprecations; ChatGPT Release Notes (help.openai.com) for ChatGPT product features and product-side model retirements; and openai.com announcements for company-level events (partnerships, supply terminations, policy). See the full table in section 2.

2. Can I still use a model via the API after it leaves ChatGPT?

Usually yes. When GPT-4o, GPT-4.1, GPT-4.1 mini, o4-mini, and GPT-5 retired from ChatGPT on 2026-02-13, the official notes stated that API availability remained unchanged. API removals follow the deprecation entries in the API changelog -- a separate timeline from product retirements.

3. How long do models stay in ChatGPT?

About 90 days by convention: the official notes on 2026-06-12 confirmed that models generally remain available in ChatGPT for roughly 90 days after a successor is released. Examples: o3 was announced 2026-05-28 and went offline 2026-08-26 (90 days); GPT-4.5 was announced 2026-05-28 and went offline 2026-06-26 (30 days). The successor's launch date starts the countdown.

4. Wasn't DALL·E already discontinued? Why another retirement?

Two layers: the DALL·E 2/3 API models were hard-removed from the API on 2026-05-12, while the official DALL·E GPT inside the ChatGPT product only retires on 2026-08-30. Always identify the layer -- the same brand can live months longer on one surface than the other.

5. How does a team operationalize this?

Five steps: fix a weekly review slot and an owner; curl the surfaces into dated HTML snapshots; grep for deprecat / retire keywords and dates; diff against last week's snapshot for increments; and convert each finding into an engineering action (update model config, schedule a migration, adjust budget). Full commands are in section 5.

6. What in company announcements should developers watch?

Model supply relationships. For example, OpenAI announced on 2026-08-28 that it will end model supply to Cursor (effective 2026-11-12) -- events like this directly affect where your models come from, with a larger blast radius than any single API change.

Next Steps

• Want a curated timeline of H2 2026 updates? Read OpenAI 2026 H2 Model and API Update Timeline.
• Want a continuously maintained model-level changelog? Read OpenAI Model Update Log (2026, Continuously Updated).
• Which currently available models should you pick? Read GPT Models Complete Guide (2026-07): Choosing Between GPT-5.6 Sol / Terra / Luna.
• What updated this week? Read OpenAI Ecosystem Week 41 Flash (2026-08-28 to 2026-08-29): Cursor Supply Ends / Multi Google Accounts / DALL·E GPT Sunset Countdown.