OpenAI's Assistants API shuts down today — and the next five shutdowns all hand work back to you
TL;DR: The Assistants API shuts down today, 26 August 2026, exactly twelve months after OpenAI announced it. That clock is fair, and the product had run since DevDay in November 2023. The story is what sits behind it on the calendar: five more shutdowns before 20 January 2027, including Agent Builder on 30 November — under fourteen months after it launched as the centrepiece of AgentKit. Look at what each replacement actually asks of you. Assistants’ stored configuration becomes “move it into your application code.” Agent Builder becomes the Agents SDK, in your repository. ChatKit’s OpenAI-hosted backend becomes ChatKit on your infrastructure. The Evals platform’s migration guide points at a third-party open-source tool. For you: this is not a stability problem, it is a direction. OpenAI is un-hosting its convenience layer, and the layer you build on next should be priced with a twelve-month assumption.
What actually shuts down, and when
Six waves in five months, all dates from OpenAI’s own deprecations page:
| Date | What stops working | Where you are sent |
|---|---|---|
| 26 Aug 2026 (today) | Assistants API (beta) | Responses API + Conversations API |
| 23 Oct 2026 | gpt-3.5-turbo, gpt-4, gpt-4-turbo, o1, o1-pro, o3-mini | Current GA aliases |
| 31 Oct 2026 | Evals platform goes read-only | — |
| 30 Nov 2026 | Agent Builder, reusable Prompts (v1/prompts), Evals dashboard + API | Agents SDK; “your application code”; Promptfoo |
| 11 Dec 2026 | gpt-5-2025-08-07, gpt-5-mini, gpt-5-nano, o3-2025-04-16, o3-pro-2025-06-10 | gpt-5.6-sol / terra / luna |
| 20 Jan 2027 | gpt-realtime, gpt-audio, gpt-4o-audio, gpt-4o-realtime | gpt-realtime-2.1, gpt-audio-1.5 |
Read that table the way a vendor would want you to and it is unremarkable housekeeping. Old models retire, betas graduate or die, and the notice periods are real: twelve months on Assistants, roughly six on Agent Builder, six months on the December snapshots.
Read it the way a buyer should and something else shows up.
The pattern is not deprecation. It is un-hosting.
Retiring a model is a cost you can plan for: swap a string, re-run your evals, ship. Retiring a hosted abstraction is different, because the work it was doing does not disappear. It moves.
Four shutdowns on that calendar, four moves in the same direction:
Assistants → your code. OpenAI stored your assistant configurations, threads and run state. The Responses and Conversations APIs keep conversation persistence, so this is a partial handback rather than a full one — but the configuration layer goes.
Reusable Prompts → your code, explicitly. This is the sharp one. Prompts was the feature that carried Assistants’ central idea — versioned, server-side configuration you point a request at — into the Responses era. It shuts down on 30 November, and OpenAI’s guidance is to move reusable prompt content into your application code. Teams that performed the recommended 2025 migration are being asked to migrate again in 2026, and the destination this time is their own repository.
Agent Builder → your code. Launched at DevDay on 6 October 2025, deprecated 3 June 2026, dead 30 November 2026. The visual canvas that composed agent workflows becomes the Agents SDK. The export button generates TypeScript or Python but does not reproduce the workflow graph, and OpenAI does not guarantee behaviour transfers unchanged.
ChatKit’s hosted backend → your infrastructure. Easy to miss, and the most expensive. The recommended integration let you embed ChatKit’s frontend while OpenAI ran the agent backend. That path ends with Agent Builder. The surviving advanced integration runs ChatKit on your own infrastructure against a backend you operate. A developer in OpenAI’s deprecation thread put the consequence plainly: for shops running many small client deployments, self-hosting each backend becomes “economically unviable.”
And the Evals platform closes with a migration guide titled for Promptfoo — an open-source tool OpenAI does not own.
None of these is a scandal. Together they are a strategy, and it is legible: OpenAI is retreating from stateful, hosted developer scaffolding back toward stateless inference primitives, and pushing state, orchestration and hosting onto the customer. That is the same bet visible in the Codex app-server repositioning — ship the loop as something you run, not something OpenAI runs for you.
The half-life problem
The models are not the churn. Set the two categories side by side:
| Launched | Shut down | Lifespan | |
|---|---|---|---|
| Assistants API | 6 Nov 2023 | 26 Aug 2026 | ~34 months |
| GPT-5 snapshots | 7 Aug 2025 | 11 Dec 2026 | ~16 months |
| Agent Builder | 6 Oct 2025 | 30 Nov 2026 | ~14 months |
Assistants got a genuinely long run and a full year of notice. The GPT-5 snapshots get about sixteen months, which is ordinary for a pinned frontier snapshot.
Agent Builder is the number that should change behaviour. It was not a beta experiment tucked in a corner — it was announced at DevDay as the centrepiece of AgentKit, positioned as the way to compose production agent workflows. It received its deprecation notice roughly eight months after launch. As one developer wrote in the thread: “Building and offering a tool that in around 6 months is deprecated…??? Not a sound business strategy.”
That is the planning input. Not “OpenAI is unreliable” — the model clocks say otherwise — but the newer and more convenient the layer, the shorter its expected life, and DevDay billing is not evidence of longevity.
How this compares to Anthropic and Google
The fair version of this comparison does not flatter one vendor.
Anthropic publishes the weakest headline notice period of the three: at least 60 days before retiring a publicly released model. It pairs that with the strongest substantive commitment anyone has made — preserving the weights of every publicly released model for at minimum the lifetime of the company. Claude Opus 3 retired on 5 January 2026 and remains reachable by paid subscribers and on the API by request. Sixty days is thin; “we will not destroy the artefact” is substantial. Anthropic is not blameless on withdrawal generally — it pulled the only tier carrying a published uptime target this month.
Google publishes shutdown dates as “the earliest possible dates on which a model might be retired” with no committed minimum window for stable models and two weeks for preview models. In practice gemini-2.0-flash ran from 5 February 2025 to 1 June 2026 — about sixteen months, matching OpenAI’s snapshot clocks. Google is also perfectly capable of hard cutovers: it shut down Gemini CLI for consumer tiers with roughly a month’s notice and no feature parity in the replacement.
OpenAI is competitive on models and distinct on tooling. It also has form on consumer-facing hard stops — Sora and DALL-E 3 both went dark on tight clocks earlier this year.
The conclusion is not that one lab is safe. It is that hosted convenience is the risky category at every lab, and OpenAI simply has the most of it exposed.
What to actually do
Inventory your pinned strings this week. Grep for every model snapshot and platform feature you call, then map each to its date from the table above. Pinning a snapshot is the responsible engineering choice, which is exactly why it quietly becomes a dated liability. Make the check recurring — a scheduled diff against the vendor deprecation pages turns every future shutdown into a ticket instead of an outage.
Do the Prompts migration once, properly. If you use v1/prompts or dashboard Prompt objects, you have until 30 November. Move prompt content into version-controlled application code rather than looking for another server-side home. That is both OpenAI’s guidance and the thing you would have wanted anyway: prompts in git are diffable, reviewable and portable.
Budget the ChatKit hosting change in money, not story points. If you were on the recommended integration, 30 November converts an OpenAI-hosted backend into infrastructure you pay for and page on. Price it now.
Keep evals outside the vendor. OpenAI closing its own Evals platform and pointing at Promptfoo is an unusually clear signal. Your evaluation suite is the asset that makes every future migration cheap; it should never live somewhere a deprecation notice can reach it. It is also what makes a second vendor a routing change rather than a project — the case for neutral gateways and model-agnostic orchestration.
Adopt the twelve-month rule. When a lab ships a hosted developer convenience, assume roughly a year of life unless it is a core inference primitive. Prototype on it freely. Do not put anything on its critical path that would be expensive to rebuild.
The verdict
Nothing on this calendar is a betrayal. Twelve months of notice on a beta is fair, six months on a snapshot is standard, and every lab retires models — Google on similar clocks, Anthropic with a shorter notice and a better preservation promise.
What the calendar does reveal is a direction that no single announcement made explicit. Across four separate shutdowns, OpenAI is systematically handing state, orchestration and hosting back to its customers and keeping the inference primitives. Whether that is a retreat from an unprofitable layer or a deliberate narrowing to what it does best matters less than the planning consequence, which is the same either way: the convenience layer is where deprecation risk concentrates, and it is compounding faster than the models are.
Budget for it the way you budget for the promotional pricing clocks and the 2027 hardware cost floor — as a dated liability with a known expiry, not as a surprise.
For the wider picture, see the best AI agent tools comparison, the best AI harnesses guide, and the best AI coding tools roundup. On the individual platforms named here: the Codex review, the ChatGPT review, the Claude review and the Gemini review.
Frequently asked questions
Does anything break today if we are still calling the Assistants API?
Yes — today is the shutdown, not the start of a grace period. OpenAI announced the deprecation on 26 August 2025 and set the sunset exactly one year later, on 26 August 2026. Once a beta endpoint reaches its shutdown date, calls to it stop being served; assistants, threads and the associated run objects are no longer addressable through that API surface. If you have production traffic still hitting it, the migration is now an incident rather than a project. The destination is the Responses API together with the Conversations API, which is where OpenAI folded the parts of Assistants worth keeping — persistent conversation state and the code interpreter tool. What does not carry over cleanly is anything that relied on server-side assistant configuration objects, which is the part developers in OpenAI's own community thread flagged as the gap between the parity claim and the migration reality. If you are reading this with live traffic on the old endpoint, pull your vector stores and file references first: those are the assets that are painful to reconstruct.
Is OpenAI actually less stable than Anthropic or Google here, or does everyone do this?
On models, everyone does this, and OpenAI's clocks are unremarkable. OpenAI gave twelve months on the Assistants API and roughly sixteen months on the GPT-5 and o3 snapshots that retire on 11 December 2026. Google shipped gemini-2.0-flash on 5 February 2025 and set its shutdown for 1 June 2026 — about sixteen months — and publishes only 'earliest possible' dates with no committed minimum window for stable models, plus two weeks for preview models. Anthropic promises at least 60 days on paper, which is the weakest headline number of the three, but pairs it with the strongest substantive commitment: preserving the weights of every publicly released model for at minimum the lifetime of the company. Claude Opus 3 was retired on 5 January 2026 and remains available to paid subscribers and on the API by request. So the fair comparison is not that OpenAI retires models faster. It is that OpenAI has a second category — hosted platform tooling — that turns over much faster than its models do, and neither rival has as much of that category exposed to customers.
We already migrated from Assistants to the Responses API with stored Prompts. Do we have to migrate again?
Partly, and this is the sharpest edge in the whole calendar. The Responses API itself is not going anywhere — that migration was sound and you should not unwind it. What expires is the reusable Prompts feature, the v1/prompts endpoint and the dashboard-created Prompt objects, which shut down on 30 November 2026 alongside Agent Builder. Reusable Prompts was the piece that reproduced Assistants' core proposition: a versioned, server-side configuration you point a request at rather than assembling in your own code. OpenAI's stated guidance is to move reusable prompt content into your application code. So a team that did exactly what it was told in 2025 gets a second migration in 2026, and the endpoint of that second migration is not another OpenAI feature — it is your repository. The practical read is that server-side stored configuration is not a stable place to keep anything on this platform, and the version-controlled alternative you would have built anyway is now also the officially recommended one.
What happens to our Agent Builder workflows on 30 November 2026?
They stop running, and the export path is lossy. Agent Builder launched at DevDay on 6 October 2025 as the centrepiece of AgentKit, was deprecated on 3 June 2026, and shuts down on 30 November 2026 — under fourteen months from flagship launch to switch-off, with the deprecation notice landing about eight months in. OpenAI's migration route is the Agents SDK: the builder has an export that generates TypeScript or Python from your canvas, but the export does not reproduce the workflow graph and does not guarantee that every behaviour transfers unchanged, so treat the generated code as a starting scaffold that needs testing rather than a finished port. There is a second cost that is easy to miss. If you used the recommended ChatKit integration, OpenAI was hosting the agent backend for you. That option ends with Agent Builder; the surviving advanced integration runs ChatKit on your own infrastructure via the Python SDK against a backend you operate. For anyone running many small client deployments, that is not a code change — it is a new hosting bill and a new operational surface, which is precisely the objection developers raised in OpenAI's own deprecation thread.
Does this change which vendor we should pick?
It should change what you build on, more than who you buy from. Choosing a lab on deprecation policy alone would be a strange trade — model quality, price and latency move far more money than migration overhead does, and OpenAI's model retirement clocks are competitive with Google's. What this calendar should change is the altitude at which you integrate. Every shutdown on the list between now and January hits a hosted convenience layer or a pinned snapshot, and none of them hits the core inference primitives: chat-style completions, the Responses API, structured outputs, function calling and the generally-available model aliases. Those primitives have been stable across the whole period, and they are also the surface that maps most cleanly onto other vendors. So the durable posture is to depend heavily on primitives, use the convenience layers for prototyping with an assumed twelve-month life, and keep your prompts, tool definitions and evaluation suite in your own repository where no vendor calendar can reach them.
What is the single highest-value thing to do this week?
Build the inventory, because most teams cannot currently answer the question. Grep your codebase and infrastructure for every pinned model snapshot string and every OpenAI platform feature you call, then put each one against the published shutdown dates: 23 October 2026 for gpt-3.5-turbo, gpt-4, gpt-4-turbo, o1, o1-pro and o3-mini; 31 October for Evals going read-only; 30 November for Agent Builder, v1/prompts and the Evals dashboard and API; 11 December for the GPT-5 and o3 snapshots; 20 January 2027 for gpt-realtime, gpt-audio, gpt-4o-audio and gpt-4o-realtime. Pinned snapshots are the ones that bite silently, because pinning is the responsible engineering choice that quietly converts into a dated liability. The second job is to make the inventory permanent rather than a one-off: a scheduled check that diffs your model strings against the vendor deprecation pages costs an afternoon and turns every future shutdown from a surprise into a ticket.
Sources
- OpenAI API — Deprecations (official list of shutdown dates)
- OpenAI Developer Community — Assistants API beta deprecation, 26 August 2026 sunset
- OpenAI Developer Community — Deprecation notice: Agent Builder
- OpenAI — Introducing AgentKit (Agent Builder launch, DevDay, 6 October 2025)
- TechCrunch — OpenAI launches API that lets developers build 'assistants' into their apps (6 November 2023)
- Anthropic — Commitments on model deprecation and preservation
- Anthropic — Model deprecations (Claude Platform Docs, 60-day notice policy)
- Google — Gemini API deprecations and shutdown dates
Related tool reviews
Questions or corrections? Email Pick Right. Want the full list? See all news.