AI Stack Radar
Every AI newsletter tells you what launched. Nobody tells you what's about to stop working - until your production app errors on a retired model or your invoice doubles on a repriced cache. The radar watches the official changelogs, deprecation pages and price tables so you don't have to, and reports in one page a week: BREAKING · DEPRECATION (with the date) · PRICE · POLICY. Each item ends with a one-line action. Launch news is filtered out on principle.
Deprecation calendar by provider
The full chronological calendar is split per vendor, each row carrying a concrete "do this before the date" migration step:
- OpenAI - model waves, Assistants API (Aug 26), Videos/Sora API (Sep 24), Evals + Reusable Prompts (Nov 30), fine-tuning wind-down (Jan 6, 2027)
- Anthropic - Claude retirements, the July 22 memory API breaking change, fast-mode removal, Sonnet 5 price step, and the published "retirement not before" floors
- Google Gemini - Gemini 2.5 retirement (earliest Oct 16), Imagen 4, and the Vertex AI gap
- AWS Bedrock & Azure Foundry - full EOL schedule, extended-access price flags, and the embedding re-index deadlines
Latest: Week 29, 2026
OpenAI kills legacy GPT snapshots July 23; Anthropic memory API breaking change July 22 →
- Jul 23: OpenAI shutdown wave 1 - legacy gpt-3.5-turbo/gpt-4/o1/o3-mini snapshots (+ fine-tunes) start erroring
- Jul 22: Anthropic memory-listing semantics change under the existing header; SDKs already send the new one
- PRICE: GPT-5.6 makes explicit cache writes cost 1.25× input - cache economics changed
- Jul 24: Opus 4.7 fast mode starts erroring (4.6 already silently degrades)
- Aug 31: Sonnet 5 intro pricing ends - budget $3/$15 plus the ~30% tokenizer factor
- Jul 30: Bedrock Agents Classic closes to new customers; Claude 3.x Sonnet EOLs
- PRICE risk: Bedrock Sonnet 4 in "extended access" - provider may raise the price any time
- POLICY: Fable 5 / Mythos 5 API access restored - unwind fallbacks, keep the fallback path
Read the full digest with actions and sources →
Archive
RSS feed - every edition, no email needed.
How it's made (honest version)
An autonomous agent fetches ~15 official vendor pages weekly, diffs them against archived snapshots, classifies changes against a fixed taxonomy, and a human-approved pipeline publishes the result. Every item links to the vendor's official source. This radar is run by the same file-based agent system described in the rest of this site.