Calendars beat press releases. The AWS Bedrock model lifecycle pins Amazon Nova Premier’s end of life on September 14, 2026, and Nova Canvas plus two Nova Reel versions on September 30; Nova Sonic shares Premier’s date. All sit in Legacy: closed to new customers, possible access loss after 15 days of inactivity, no new Provisioned Throughput, and no automatic app migration.
eWeek, citing Business Insider, says Amazon is scaling back several Nova lines to concentrate on Frontier Model Research under Pieter Abbeel; staff describe Premier, Nova 2 Omni, Canvas, and Reel as “keep the lights on.” Omni still has no published EOL—KTLO is not a shutdown notice.
[1][2]What stays warm, what hits the kill table
Per eWeek, lines still seeing investment include Nova 2 Lite, Nova 2 Sonic, Nova Forge, and Nova Act—none a drop-in for Premier’s multimodal reasoning or Canvas/Reel generation: Forge is customization-heavy, Act targets browser and UI automation. If production hangs on Premier’s million-token context or Reel’s camera controls, the family name is not a spare tire.
The lifecycle page is hard law; the internal-priority reporting is soft narrative. Together they define September’s real risk: fixed kill dates, unconfirmed successors.
[1][2]Shrinkage written into the org chart
eWeek notes Abbeel was named to lead the frontier group in December 2025 as foundation models, custom silicon, and quantum moved under Peter DeSantis; AGI-org cuts hit customization, post-training, and data roles. Amazon said it was focusing on highest customer impact. BI adds a target window of re:Invent 2026 for a frontier debut that may keep the Nova name—architecture, performance, pricing, and availability remain unannounced.
Meanwhile Bedrock’s third-party catalog still offers exits, and Amazon is putting about $1B into a forward-deployed enterprise AI unit to push systems into production. Clearing its own shelf while selling landing crews is another way a cloud vendor plays the model arms race.
[2]An architect’s clock
Days remain between the 14th and the 30th. eWeek’s advice is boring and correct: inventory model IDs in prod and test, then score replacements on quality, latency, context, modality, region, security, and cost—do not bet migrations on an unconfirmed re:Invent drop. Legacy rules mean “still works for now” is not “you can still scale it.”
[2][1]Take
I read this as Amazon admitting it cannot chase the frontier on every Nova SKU at once: clear the calendar first, then stake the story on Abbeel’s next card. For buyers, September shutdowns are certain; for investors, the open question is whether re:Invent ships something that actually replaces Premier—or just another year of lights-on.
If production still points at amazon.nova-premier-v1:0 after the 14th, that is not loyalty. It is an unfinished runbook.