Deadlines
Every end date a provider has announced and this desk is tracking, counting down. Windows closing, prices changing, models retiring.
Only published dates appear here. Nothing inferred from a pattern, nothing expected-but-unannounced, nothing from rumour — if a provider has not said it, it is not on this page. These are records of what was announced, not predictions. Providers move their own dates and are entitled to; when that happens the change is itself a record and gets filed.
Put these dates in your own calendar
A live subscription, not a download — when a provider moves a date, the event moves with it. The 17 rows that need something from you carry a reminder seven days out; the rest arrive as quiet all-day entries.
Paste https://hugin.studio/deadlines.ics into your calendar app’s “subscribe by URL”. Works in Apple Calendar, Google Calendar, Outlook and anything else that speaks iCalendar.
These need something from you
Most deadlines are worth knowing and nothing more. These are the exceptions — the ones where waiting costs you something.
- in 3 daysOpenAIRetirement
The Videos API and the entire Sora 2 model family shut down
Before this date: Decide what replaces it, not which model to switch to. This is a vendor evaluation with a date on it — new contract, new prices, a different safety posture — so it needs whoever can authorise that, not only whoever owns the code.
Nothing to move toThe ledger's replacement column reads "---" for all six rows, and no migration sentence appears anywhere in the announcement. Measured 2026-09-07: these are the only six rows on the page in that condition, out of 49 with a replacement column.
Covers the Videos API plus sora-2, sora-2-pro and three dated snapshots. The consumer app went earlier, on April 26, 2026; the help-centre article for it commits to deleting Sora content after a final export window it describes only conditionally.
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
- in 7 daysOpenAIRetirement
Legacy snapshots gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106 shut down
Before this date: Move any pinned legacy snapshot to the named substitute, gpt-5.6-terra, before the date.
Moves togpt-5.6-terra, named for all four ids.
Part of a stated multi-wave retirement of older models with declining usage. A substitute model is named for every id in this group.
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
- in 9 daysGoogleRetirement
The gemini-omni-flash-preview endpoint is deprecated in the Gemini API
Before this date: Search your configuration for the exact string gemini-omni-flash-preview and repoint it at gemini-omni-1.1-flash.
Moves togemini-omni-1.1-flash, released August 27, 2026 as the GA version of the same model.
Google's sentence says the endpoint 'will be deprecated', not removed, and names the GA model in the neighbouring entry rather than as a stated replacement. The changelog redirects an unauthenticated plain client to a sign-in page, so this desk reads it through a rendering fetcher.
Operative page re-read: Read, and it still says what this row says: 4 recorded terms present.
- in 10 daysOpenAIRetirement
gpt-5.4-cyber is removed from the API
Before this date: Repoint any pinned gpt-5.4-cyber model id at gpt-5.6-cyber before the date.
Moves togpt-5.6-cyber, named in both the announcement sentence and the ledger's replacement column.
Announced in a ledger section dated September 11, 2026 — twenty days of notice, the shortest on the ledger's pending schedule. The ledger records a schedule; it keeps carrying ids after their date, so it cannot by itself confirm a shutdown happened.
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
- in 11 daysGitHubRetirement
GitHub Copilot deprecates Claude Opus 4.7, Gemini 3.5 Flash, Gemini 3.6 Flash and Kimi K2.7 Code across all Copilot experiences
Before this date: Copilot Business and Enterprise administrators: enable the alternative models in Copilot model policies before the date — GitHub's notice says they may need to, and until they do users lose the old model without getting the new one.
Moves toClaude Opus 5 for Opus 4.7, Gemini 3.8 Flash for both Gemini Flash models, Kimi K3 for Kimi K2.7 Code — GitHub's suggested alternatives.
A retirement from ONE tool, not from the model's maker: Anthropic's deprecations page lists claude-opus-4-7 as Active with a retirement not sooner than April 16, 2027, and Google lists no date for its Gemini 3.x Flash models. The models remain available elsewhere; they leave Copilot.
Operative page re-read: Read, and it still says what this row says: 3 recorded terms present.
- in 11 daysGoogleRetirement
gemini-2.5-flash-image reaches its earliest listed shutdown date in the Gemini API
Before this date: Move calls to gemini-3.1-flash-image-preview before the date. Google lists October 2 as the earliest possible shutdown, so the model may keep answering past it — plan for the date, do not wait for a failure.
Moves togemini-3.1-flash-image-preview, named in the deprecations table's replacement column.
Google's table states listed shutdown dates are 'the earliest possible dates on which a model might be retired' — a floor, not a schedule. The page redirects an unauthenticated plain client to a sign-in flow, so this desk reads it through a rendering fetcher.
Operative page re-read: Read, and it still says what this row says: 8 recorded terms present.
- in 32 daysOpenAIRetirement
A second legacy wave shuts down seventeen listed ids: the gpt-3.5/GPT-4-era snapshots (gpt-3.5-turbo-0125, gpt-4-0613, gpt-4-1106-preview, gpt-4-turbo, gpt-4.1-nano, gpt-4o-2024-05-13), the reasoning models o1, o1-pro, o3-mini and o4-mini, the original gpt-image-1 — and six fine-tuned lines including ft-gpt-4, ft-babbage-002 and ft-davinci-002
Before this date: Substitutes are named per id — gpt-5.6-terra for the 3.5 and o4-mini lines, gpt-5.6-sol for the GPT-4 and o1/o3 lines, gpt-5.6-luna for the nano line, gpt-image-2 for gpt-image-1. Check which of your pins is in which wave; the September and October groups do not share a date.
Moves toPer id — gpt-5.6-terra for the 3.5 and o4-mini lines, gpt-5.6-sol for the GPT-4 and o1/o3 lines, gpt-5.6-luna for gpt-4.1-nano, gpt-image-2 for gpt-image-1.
A separate group from the September wave, announced on its own date. The ledger lists each snapshot alias alongside the model name, so a pin on either form is covered. Until August 26 this row described only six of the seventeen ids; the fuller list above is what the cited section has carried since at least August 10.
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
- in 40 daysOpenAIRule due
Existing evals on the deprecated Evals platform become read-only
Before this date: Export or migrate any eval you still need to edit before this date — after it, existing evals can be read but not changed, a month before the platform shuts down.
The first of two dated steps for the Evals platform — the shutdown itself is a separate row on November 30. Announced June 3, 2026; the ledger names Promptfoo as a migration path.
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
- in 42 daysxAIRetirement
The grok-imagine-image-quality model name is retired from the xAI API — and keeps answering, served by a different model
Before this date: Call grok-imagine-image-2.0 explicitly before November 2 and choose the quality you pay for; otherwise the retired name keeps working but is served by a different model at low quality.
Moves togrok-imagine-image-2.0 at low quality, which the retired name silently resolves to after the date.
Nothing breaks on the date: 'The slug continues to resolve, so you do not need to change your code to avoid breakage.' What changes is the model behind the name and its price, which xAI states is $0.01 per image lower at every resolution.
Operative page re-read: Read, and it still says what this row says: 3 recorded terms present.
- in 70 daysOpenAIRetirement
The v1/prompts API and reusable prompt objects shut down
Before this date: Move reusable prompt content into your application code — the ledger's stated migration is inlining, not a replacement API.
Moves toNo replacement API. The page's stated migration is to "move reusable prompt content into your application code", with a linked Migrate from prompt objects guide — a path, and not a product.
Announced June 3, 2026, with prompt creation de-emphasized in the platform the same day. Lands on the same date as the Agent Builder and Evals shutdowns — three products end on November 30.
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
- in 70 daysOpenAIRetirement
The Evals dashboard and API shut down
Before this date: Migrate eval workflows before the date; the ledger names Promptfoo as the documented path. Evals are already read-only from October 31.
Moves toPromptfoo, a third-party tool: the page links "Moving from OpenAI Evals to Promptfoo for a migration path". Worth reading twice — the continuation is off-platform.
The second Evals step — the October 31 read-only cliff is its own row. Graders documented for eval workflows are stated to be part of this transition; fine-tuning timelines are separate.
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
- in 70 daysOpenAIRetirement
Agent Builder shuts down
Before this date: Migrate to the Agents SDK or ChatGPT Workspace Agents — both are named on the page as the continuation path.
Moves toThe Agents SDK or ChatGPT Workspace Agents, via a linked Migrate from Agent Builder guide; ChatKit is stated to remain available.
The ledger states that ChatKit remains available. The shutdown is of the builder product, not of every agent surface.
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
- in 71 daysOpenAIRetirement
The older GPT Image models — gpt-image-1-mini, gpt-image-1.5 and the chatgpt-image-latest alias — are removed from the API
Before this date: Repoint image calls at gpt-image-2, the named replacement for all three.
Moves togpt-image-2, named for all three ids.
chatgpt-image-latest is a floating alias, so nothing in your code names a version that could warn you — the id keeps working right up to the date and then does not. Notified June 2, 2026.
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
- in 81 daysOpenAIRetirement
The first-generation GPT-5 snapshots — gpt-5, gpt-5-mini, gpt-5-nano and gpt-5-pro — and the o3 and o3-pro reasoning models are removed from the API
Before this date: Repoint at the named substitutes: gpt-5.6-sol for gpt-5, gpt-5-pro, o3 and o3-pro (pro via reasoning.mode), gpt-5.6-terra for gpt-5-mini, gpt-5.6-luna for gpt-5-nano.
Moves togpt-5.6-sol, -terra or -luna per snapshot; the two pro models map to gpt-5.6-sol with reasoning.mode: pro.
Announced June 11, 2026. Sixteen months after GPT-5's release, its launch snapshots retire — every id in this group is snapshot-dated, so a pinned snapshot fails on the date while a floating alias will already have moved.
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
- in 107 daysOpenAIRule due
Active existing customers can no longer create new self-serve fine-tuning jobs
Before this date: If a fine-tuned model is part of your roadmap, the job has to exist before this date — creation ends, not inference.
Scoped to CREATING jobs. The ledger states that inference on already fine-tuned models is disabled only when the underlying base model is deprecated, which is a separate date per base model.
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
- in 121 daysOpenAIRetirement
Every legacy audio, realtime and transcription family is removed from the API — gpt-realtime, gpt-audio, gpt-4o-audio, gpt-4o-realtime, their mini variants, and gpt-4o-mini-transcribe
Before this date: Repoint at the named replacements: gpt-realtime-2.1 and gpt-realtime-2.1-mini for realtime, gpt-audio-1.5 for every audio family, gpt-4o-mini-transcribe-2025-12-15 for transcription.
Moves togpt-realtime-2.1 and gpt-realtime-2.1-mini for realtime, gpt-audio-1.5 for the audio families, gpt-4o-mini-transcribe-2025-12-15 for transcription.
Announced July 20, 2026. This group retires whole FAMILIES, not dated snapshots — a floating family id like gpt-realtime keeps working right up to the date with nothing in your code naming a version that could warn you, the chatgpt-image-latest trap at family scale.
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
- in 158 daysOpenAIRetirement
The legacy transcription models are removed from the API — whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize
Before this date: Repoint at gpt-live-transcribe or gpt-transcribe, the two replacements the ledger names for all four ids.
Moves togpt-live-transcribe or gpt-transcribe, named for all four ids.
Announced August 26, 2026, and the farthest date on the ledger — which is exactly why it is the one to write down. whisper-1 is embedded in a great deal of shipped code as a default rather than a choice, so this is more likely to be found in a dependency than in a config.
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
Everything else upcoming 0
Nothing here right now: all 17 upcoming dates need something from you, and they are listed above. This section exists for the dates that are worth knowing and nothing more.
Passed 9
Kept rather than deleted, and each one now says what happened — or admits that nothing was filed. A date that has gone by is still the record of what was announced, and quietly dropping it would make this page look more prescient than it is.
Show the 9 that have passed
- 8 days agoAnthropicPromotional window ends
The 50% Claude Code weekly-limit boost ends
Confirmed by this desk's own re-readEnded at its stated 11:59 PM PT on September 13, when the terms article was rewritten — and limits did not fall back: 'starting September 14, 2026, weekly limits in Claude Code are 25% higher than they were before the promotion for Pro, Max, Team, and seat-based Enterprise plans.'
Read the entry that records itExtended three times after its first stated end (July 13, then August 19, August 31 and September 13); the terms article is the page that moved each time. The stretch between the first window's stated July 13 end and the July 18 renewal has no provider-authored record and is carried at coverage tier only.
Operative page changed: Read, and something changed: 3 recorded terms no longer on the page. Open the page before relying on this row.
- 21 days agoOpenAIRetirement
GPT-5.4 and GPT-5.4 mini stop being available in Codex for users signed in with ChatGPT
Before this date: If you use those models in Codex via ChatGPT sign-in, switch to a recommended replacement or authenticate with an API key.
Moves toA recommended replacement model in Codex, or the same models via API-key authentication — the retirement is scoped to an access path, not to the model.
Confirmed by this desk's own re-readThe page still describes it as upcoming. Read 2026-09-14: the Codex models page, since reworded, says 'GPT-5.4 and GPT-5.4 mini retire from Codex on August 31, 2026' and still offers the codex -m gpt-5.4 and codex -m gpt-5.4-mini commands. The page was not updated; whether ChatGPT sign-in access actually ended is not something it settles.
Read the entry that records itScoped to an access path, not the models: both remain on the OpenAI API and in Codex sessions authenticated with an API key.
Operative page changed: Read, and something changed: 1 term still listed after the date it was due to go. Open the page before relying on this row.
- 21 days agoAnthropicPrice change
Sonnet 5 introductory API pricing ($2/$10) ends
Confirmed by this desk's own re-readReversed: the introductory $2/$10 rate 'is now the standard price', and the page says the 'previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur.' First found on September 1; unchanged on 2026-09-14.
Read the entry that records itStandard pricing of $3/$15 is stated for September 1. These are the provider's published terms, not a quote for any account.
Operative page re-read: Read, and it still says what this row says: 2 recorded terms present.
- 22 days agoOpenAIRetirement
The official DALL·E GPT is retired in ChatGPT
Before this date: Download any images you want to keep before this date — the notice says so explicitly.
Moves toImage generation in ChatGPT itself; the notice states user-created GPTs with image generation enabled are unaffected.
Not establishable by this deskNot establishable. The ChatGPT release notes, readable on 2026-09-14 after refusing this desk's client through September 1, still carry only the notice 'On August 30, 2026, we're retiring the official DALL·E GPT in ChatGPT'; nothing after it records the retirement.
Read the entry that records itUser-created GPTs with image generation enabled are stated to be unaffected. ChatGPT Images remains available for creating and editing.
Operative page re-read: Read, and it still says what this row says: 2 recorded terms present.
- 26 days agoOpenAIRetirement
The Assistants API shuts down on two vendors at once — OpenAI names the Responses and Conversations APIs as the replacement, Microsoft retires Azure OpenAI Assistants the same day and names Microsoft Foundry Agents instead
Before this date: Migrate before the date, but check which vendor you are actually on first: direct OpenAI integrations go to the Responses or Conversations API, Azure OpenAI deployments go to Microsoft Foundry Agents. The two notices name different destinations and neither mentions the other.
Moves toThe Responses or Conversations API for direct OpenAI integrations; Azure customers are pointed at Microsoft's own path, which is a different vendor's recommendation.
Confirmed by this desk's own re-readShut down, and establishable from outside: unauthenticated, /v1/assistants returns 404 with an empty body — the same answer this API gives for a route that does not exist — while a live route still answers 401 with a JSON body. Six spaced probes across 2026-08-26 and 2026-08-27, each run with the missing-route control alongside, never moved. Microsoft's Foundry (classic) notice, future-tense on the morning of the 26th, now reads "The Assistants API is retired."
Read the entry that records itAn API shutdown affects integrations rather than chat users, and both replacements are their vendor's own recommendation rather than this desk's advice. Added 2026-08-19: this row previously carried only the OpenAI exit, which was accurate for OpenAI's surface and silent about Azure. Microsoft's Foundry (classic) Assistants page, read the same day, states the feature "is deprecated and will be retired on August 26, 2026" and directs workloads to Microsoft Foundry Agents — https://learn.microsoft.com/en-us/azure/ai-foundry/openai/concepts/assistants?view=foundry-classic
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
- 26 days agoOpenAIRetirement
OpenAI o3 is retired from ChatGPT after a stated 90-day sunset
Moves toGPT-5.6 in ChatGPT — the release notes retire o3 into the current default rather than naming a per-model substitute.
Not establishable by this deskNot establishable. Read 2026-09-14, nineteen days after the date: the model release notes still say 'OpenAI o3 will be retired from ChatGPT on August 26, 2026', in the future tense, and no later entry records the retirement. A schedule nobody updated cannot show whether it was kept.
Read the entry that records itA retirement notice records that access is ending. It says nothing about the model's quality or about why.
Operative page re-read: Read, and it still says what this row says: 1 recorded term present.
- 33 days agoAnthropicPromotional window ends
Doubled Cowork 5-hour usage limits end and standard limits resume
Not establishable by this deskNot establishable. The terms article that carried the window has answered HTTP 404 since August 20 and still did on 2026-09-14; no replacement terms article was found, and the nearest help article describes how Cowork uses limits without stating them.
Read the entry that records itA promotional window ending is a return to standard terms, not a reduction of anything anyone was entitled to. The launch post still carries the superseded August 5 date; the terms article is the one that gets edited.
Cannot be checked from here: Not checkable from here: this page refuses or challenges the identifying client this desk publishes.
- 42 days agoOpenAIRetirement
gpt-5.2-chat-latest and gpt-5.3-chat-latest shut down on the API, replaced by gpt-5.6-sol
Before this date: If your code pins either alias, move it to a supported model before the date.
Moves togpt-5.6-sol, named in the ledger's replacement column for both aliases.
Confirmed by this desk's own re-readExecuted on schedule, 94 days after the May 8 notice. The ledger still carries both ids against "Aug 10, 2026" with gpt-5.6-sol as the replacement; neither appears on the live models list, and no "-chat-latest" alias does.
Read the entry that records itScoped to the two dated chat-latest aliases. The deprecations ledger is the provider's own schedule and is where any change to this date would appear.
Operative page re-read: Read, and it still says what this row says: 31 recorded terms present.
- 43 days agoOpenAIRetirement
The Atlas browser stops working
Moves toChatGPT for browser-based agentic work, named in the article's own title.
Not establishable by this deskReported to have happened, but not establishable from any OpenAI page this desk can read: help.openai.com, openai.com and chatgpt.com all refuse or challenge the identifying client, while the developer docs serve it the whole document.
Read the entry that records itThe release notes say Atlas is scheduled to stop working on this date. A scheduled shutdown is a stated plan, not a guarantee of the hour.
Operative page re-read: Read, and it still says what this row says: 1 recorded term present.