Google silently removed the Gemini 2.5 shutdown date from its deprecation page

The date that disappeared

On July 28, 2026, Google's model deprecation page listed October 16, 2026 as the shutdown date for gemini-2.5-pro, gemini-2.5-flash and gemini-2.5-flash-lite. Six days later the date was gone. According to a dev.to analysis by the operator of an independent deprecation tracker, all three models now read "No shutdown date announced," and the October 16 date appears nowhere on the page.

The edit came with no changelog entry, no banner and no email, the author reports. It was caught at all because the tracker runs a crawler over the deprecation pages of 15 model vendors daily and archives what those pages overwrite. The consequence is concrete: a team that read the page in late July would have planned a migration against an October deadline, while a team that read it a week later would have seen no deadline at all. Both relied on the vendor's own documentation, and neither had any way to know the other reading had ever existed.

Sixty-two shutdowns in ninety days

The same crawl, snapshotted on September 24, 2026, counts 62 models scheduled for removal between that date and December 22: 12 in the final week of September, 28 in October, 12 in November and 10 in the first three weeks of December. These are the tracker's own figures, not vendor announcements.

Two details in the author's data matter more than the raw count: whether a shutdown entry names a successor model, and whether it states when the shutdown was announced.

Vendor	Shutdowns	Names a successor	States announcement date
OpenAI	36	31	1
Azure	13	10	0
AWS	6	0	6
Google	6	6	0
xAI	1	1	0

Migration planning is really two questions: what to move to, and how much notice you have. OpenAI, Azure and Google answer the first well — 47 of their combined 55 shutdowns name a replacement — but almost never say when the clock started. AWS is the mirror image: all six of its entries carry an announcement date and none names a successor. Across all 62 entries, exactly one record carries both facts. It is gpt-5.4-cyber, announced September 11, 2026, removed October 1, with gpt-5.6-cyber named as the target: a runway of twenty days.

For comparison, the author points to stated policies. Anthropic publishes a 60-day minimum notice period and AWS publishes six months, and the same Claude model was measured at 60 days of notice on one platform and 184 on the other. No published policy stands behind the twenty-day figure; it is simply what the page said that week. The tracker also declines to fill gaps with its own data — when a vendor prints no announcement date, the field stays empty rather than being backfilled with the crawler's timestamp, since polling latency is not a vendor commitment.

The deprecations no calendar can show

The tracker's archive holds a further 34 deprecations that were announced without any shutdown date. Those models are formally deprecated, have no announced end, and will never appear in a quarterly count until someone assigns them a date — possibly on very short notice.

Dates landing now

Four shutdown dates fall within a week of the analysis. Sora-2 and sora-2-pro, plus three dated variants, were set for removal on September 24 with no successor named. Babbage-002, davinci-002, gpt-3.5-turbo-1106 and gpt-3.5-turbo-instruct follow on September 28. AWS retires Nova Canvas and Nova Reel on September 30, and Azure retires gpt-4o and o3-mini on October 1.

What the author recommends

Two practices would have caught the Google reversal. First, pin the dated model string in code rather than the alias: many entries are dated variants, the variants do not all retire on the same day, and an undated alias silently remaps to whatever the vendor last pointed it at. Second, snapshot the deprecation page you are planning against — a scheduled job that fetches the page and commits the HTML to a repository produces a diff and a date, which is exactly the evidence the live Google page no longer provides.

Why it matters

A documentation page that stores current state rather than history can be correct every single day and still never tell you it changed. Silent edits make vendor documentation unverifiable at the moment it matters most, when a team is committing to a migration schedule. The tracker's numbers also show that notice periods are largely an accident of page content rather than a commitment: only one of 62 entries carries both a successor and an announcement date, and that single entry offered the shortest runway of the quarter. Until vendors treat deprecation pages as append-only records, or publish enforceable notice policies the way Anthropic and AWS do, the burden of proving what the page said — and when — falls on the engineering team.