The Week Ahead in DeAI: Sept 15–21, 2026

The week of September 15–21, 2026 opens with the largest dated deprecation cluster of the quarter: Azure AI Foundry hosted model IDs begin retiring tomorrow, September 15, per third-party deprecation trackers, with OpenAI's Videos API and Sora-2 models following on September 24.

Key facts

• September 15–24: Azure AI Foundry hosted model IDs — including gpt-realtime-mini@2025-10-06 and gpt-chat-latest@2026-06-24 — are scheduled to retire, per the llmlatency.dev deprecation tracker. Confirm each date against Microsoft's official deprecation docs before migrating.
• September 24: OpenAI's Videos API plus the sora-2 and sora-2-pro models are scheduled to retire, same source.
• September 28: legacy OpenAI instruct models gpt-3.5-turbo-instruct, babbage-002, and davinci-002 are scheduled to retire, same source.
• Today: DeepSeek V4 Pro stays online — DeepSeek's pricing page confirms the planned September 14 phase-out was withdrawn, billing unchanged.
• This week: Anthropic's Enterprise Frontier Safeguards terms are the document to watch after this morning's Fable 5.1 launch.

The deprecation cluster

Three dated retirement windows land inside the next fourteen days, all surfaced by third-party deprecation trackers and all worth verifying against the providers' own deprecation documentation before you touch production routing.

Azure's window opens first: hosted model IDs on AI Foundry — including the gpt-realtime-mini@2025-10-06 and gpt-chat-latest@2026-06-24 snapshots — are listed for retirement across September 15–24. Azure snapshot IDs are date-pinned by design, so a retirement there means a hard model-string change, not a silent upgrade path.

OpenAI's cluster follows: the Videos API and both Sora-2 model IDs on September 24, then the legacy instruct trio — gpt-3.5-turbo-instruct, babbage-002, davinci-002 — on September 28. The instruct retirements are the ones most likely to bite older completions-API integrations that never migrated to chat completions.

This is the scheduled counterpart to the silent-swap beat: dated, announced, and survivable — but only if the migration happens before the date, not after. Our migration guides cover the base-URL-and-slug mechanics.

What else is on the board

Anthropic EFS terms. This morning's Fable 5.1 / Mythos 5.1 launch came with Enterprise Frontier Safeguards, which Anthropic describes as "complete privacy (the same as a zero data retention policy)" with customer data in customer-controlled infrastructure. Phased rollout starts this fall. The claim is Anthropic's own; the published EFS terms — specifically abuse-logging carve-outs and any third-party audit commitment — are what would make it checkable. That documentation drop is the single most important privacy-beat event to watch this week.

Price Index, second observation. This morning's delta report records the week's list-price moves: DeepSeek V4 Flash 0731 down 57% on OpenRouter, GLM-5.3 Flash doubled, Kimi K3 down ~11%, and the new deepseek-v4.1-flash row matching DeepSeek's official peak rate. The next observation is scheduled for Monday, September 21.

Retention-policy divergence. Anthropic's enterprise tier is moving toward ZDR-equivalence in the same week its consumer terms introduce five-year retention for opted-in training data, while OpenAI's updated US privacy policy formalizes 30-day deletion. The enterprise/consumer split is becoming the axis procurement teams will ask about — expect more provider positioning here, and track it on the Trust Tracker.

A note on this week's inputs

The Sunday Grok week-ahead sweep that normally seeds this brief did not produce a file this week — data/discovery/2026-09-13_week_ahead.json is missing, and this morning's X sweep returned empty on a circuit-breaker failure in the grok-oauth path. This brief is therefore built from today's triaged discovery (RSS + web passes) and the dated deprecation tracker. If a calendar item feels absent, that is why; the gap is logged, not papered over.

FAQ

Which AI APIs are retiring in September 2026?

Per third-party deprecation trackers: Azure AI Foundry hosted model IDs (including gpt-realtime-mini and gpt-chat-latest snapshots) retire September 15–24, OpenAI's Videos API plus sora-2 and sora-2-pro retire September 24, and gpt-3.5-turbo-instruct, babbage-002, and davinci-002 retire September 28. Confirm each date against official deprecation docs before migrating.

Is DeepSeek V4 Pro still available after September 14, 2026?

Yes. DeepSeek's pricing page states it decided to continue providing V4 Pro API services after September 14, 2026, with billing unchanged at $1.32/$3.96 per million tokens peak — reversing an earlier phase-out plan.

What should builders watch from Anthropic this week?

The Enterprise Frontier Safeguards documentation. Anthropic launched Fable 5.1 on September 14 claiming EFS delivers privacy "the same as a zero data retention policy" — the published EFS terms, abuse-logging detail, and any audit commitment are what turn that claim into something checkable.