Quick answer: OpenAI maintains a public deprecation schedule, and the 2026 entries are no longer model-only. The Assistants API shuts down in August 2026, with the Responses API as the migration target; the Evals platform and the hosted Agent Builder surface retire in November; the legacy GPT-3.5 and GPT-4 snapshots disappear in October; and self-serve fine-tuning has stopped accepting new-job creation for new organizations, with existing customers locked out of new training jobs starting January 2027. The thing that is NOT deprecated is Chat Completions — OpenAI's deprecation page does not list it, and its own entries recommend /v1/chat/completions as the replacement for retiring edit endpoints. If your integration targets a named surface, the runbook is a scheduled swap; if it targets Chat Completions only, you are scheduling a model-family bump, not an endpoint migration.
Vendor deprecations land in three severities, and the OpenAI log needs the same sort the Twilio log does: surface-level sunsets get a work-stream, model retirements get a maintenance line item, and policy restrictions get a monitoring entry. This post sorts the schedule, names the integration shapes each class actually hits, and sketches the migration take in two bullets — where an AI-voice or AI-agent console consumes the LLM, and how the shipped model-gateway pattern reads a deprecation calendar. It is an industry explainer, not an Orbit product announcement.
What OpenAI actually deprecated — and what it did not
The public deprecation page mixes surface shutdowns and model retirements in one timeline; treating them as one class is what makes a migration budget wrong. The current entries, sorted:
- The Assistants API — surface sunset. The whole assistants surface (threads, runs, vector stores managed server-side) shuts down in August 2026. The migration target is the Responses API paired with the Conversations API — a re-authored stateful-conversation flow, not an SDK bump.
- The Evals platform — surface sunset. The hosted evaluation surface goes read-only in October 2026 and shuts down at end of November 2026, with a third-party evaluation framework (OpenAI points at Promptfoo) as the recommended target. Anything whose "quality gate" stored eval definitions in OpenAI's hosted Evals loses its test harness.
- Agent Builder — surface sunset. The hosted visual agent-builder surface retires at end of November 2026 (the ChatKit surface stays; the Agents SDK is the migration target). Prototypes built on the builder need a re-author, not a re-point.
- Legacy model snapshots — model retirements. The older GPT families (GPT-3.5-turbo, GPT-4, o1, o3-mini, o4-mini snapshots) shut down in October 2026, with image-model and additional reasoning-snapshot retirements rolling into December, and audio/realtime families into early 2027. A pinned
model: "gpt-4"string is a termination date, not a version preference. - Fine-tuning — policy restriction for new jobs. Creating fine-tuning jobs closed to new organizations in May 2026, and closes to existing customers in January 2027; inference on already-trained models continues until the base model retires. A product that "owns" its data through a scheduled nightly fine-tune has been on borrowed time since the announcement.
- Chat Completions — explicitly not deprecated. The endpoint does not appear on the deprecation schedule; the schedule's own migration rows still recommend
/v1/chat/completionsas the replacement for the retiring edit endpoints. The framing "OpenAI is deprecating Chat Completions" is wrong and makes migration budgets wrong — Chat Completions is the stable primitive; the surface-level entries above are the volatility.
The budgeting consequence: a surface sunset is a "re-author the integration" event (work-stream), a model retirement is a "bump the model string, rerun the evals" event (maintenance), and the fine-tuning restriction is a "watch the inference window" event (monitoring). Mis-sorted, the same calendar reads as one undifferentiated pile; sorted, it maps onto three different engineering cadences.
Who each class actually hits
- Assistants API retirement hits any integration holding conversation state in OpenAI's threads/runs objects rather than in its own store — demo-state agents, prototype chatbots, and counter-intuitively the integrations that were careful to keep thread ids durable in their own database. The fix is re-authoring the stateful flow against Responses + Conversations, not bumping a dependency.
- Hosted-Evals retirement hits teams whose quality gate lives inside OpenAI's hosted evaluation product — the harness evaporates. A gate re-implemented against saved failing conversations survives; a gate pointing at the vendor platform does not.
- Agent Builder retirement hits the prototypes that routed through the hosted builder's visual flow. The Agents SDK path is an ordinary application build, so the re-author is scoped — but a builder-issued artifact goes to zero without it.
- Legacy-model retirements hit any hard-pinned model string —
gpt-4,gpt-3.5-turbo, the o1/o3/o4 snapshots. For a voice pipeline the failure is first-order: a terminated model id returns an error at call time, and the call dies between the STT and the TTS. - The fine-tuning restriction hits tenant-managed custom-model strategies and "own the weights" procurement claims; inference breathing room ends at the base model's own retirement date.
The common mis-sort: teams hear "OpenAI deprecates Chat Completions" and budget a whole endpoint migration while the actually-dated event — an Assistants surface or a legacy snapshot — sits unscoped in the same log. Get the sort right before the runbook is written.
The migration take: how an AI-voice or AI-agent console should read it
- Bullet one — locate the LLM consumption point, then map the dated entries to it. A console consumes the LLM at exactly one point in the pipeline: after transcript capture and grounding, before TTS render, with tool calls dispatched into the platform. Orbit's own e2e shape runs over the MCP server surface, with the step order described in the LLM-orchestration explainer, and its quality gates authored against saved failing conversations per the AI-agent evaluation framework. Where your console exposes that same choke-point, the deprecation calendar maps surface-by-surface: an Assistants dependency is scoped at the orchestration layer, a hosted-Evals dependency at the evaluation harness, and a legacy snapshot at the model-preset layer. Scope decides the cadence.
- Bullet two — keep the model swappable; the calendar is a schedule, not an emergency. The durable answer to any LLM vendor's rotation is a tenant-owned model preset and an evaluation suite owned by the platform, not the vendor's. When the deprecation log says a snapshot family retires on a date, "swap the preset" is a maintenance line item only if the model selection is actually a preset; it becomes an architecture discovery if the pipeline hard-pins per call. The deeper runbook, including the STT/LLM/TTS time-and-cost budget the model latency feeds, is the cost-component walkthrough and the LLM spend governor explainer. A model swap is an ordinary config move on that shape — as it should be.
What this does not say (disclaimers)
This is an industry explainer, not a Devotel Orbit product announcement. Nothing in this post announces a shipped Orbit capability, a licensing change, or a pricing change: if you do not depend on OpenAI's retiring surfaces, no action is needed, and if you do, the work-stream belongs to the named surface. Orbit ships its own RAG-grounded agent stack, but a vendor-neutral stance toward the model layer is the point of this post — the OpenAI deprecation calendar is industry news a CPaaS buyer reads for risk, and the reader's integration inventory, not this blog, decides what the runbook covers. Compliance topics map to tenant-owned controls only.
Frequently asked questions
Did OpenAI deprecate the Chat Completions API?
No. OpenAI's deprecation schedule does not list Chat Completions, and the schedule's own migration rows recommend /v1/chat/completions as the replacement for the retiring edit endpoints. The surface retirements — Assistants, hosted Evals, Agent Builder — are the dated events; Chat Completions remains the recommended primitive.
What is actually retiring in the 2026 cycle?
The Assistants API (August 2026, migrate to the Responses API paired with the Conversations API), the legacy GPT-3.5/GPT-4/o1/o3/o4 model snapshots (October–December 2026), the hosted Evals platform (read-only October, shutdown end of November 2026), the hosted Agent Builder (end of November 2026, with ChatKit and the Agents SDK as the migration path), and new fine-tuning job creation (closed to new organizations May 2026; closed to existing customers January 2027).
How bad is a legacy model retirement for a voice pipeline?
First-order bad at the choke-point: a hard-pinned snapshot id errors at generation time, so the voice pipeline fails between transcript capture and TTS render. The fix is a tenant-owned model preset, so a retirement becomes a config bump plus an evaluation re-run, not a call-path rewrite.
What does the fine-tuning restriction change?
New self-serve fine-tuning jobs are closed, but inference on already-trained models continues until the base model retires. A product built on owning the data through weekly fine-tunes is on a deprecated footing; a tenant-owned RAG-grounding pipeline over a platform evaluation gate is the durable alternative.
How does an AI-agent console use the LLM?
At one choke-point in the voice pipeline — after transcript capture and grounding, before TTS render — with tool dispatch over the platform's MCP surface. The step order maps to the LLM-orchestration explainer, the e2e surface to the MCP server explainer, and the quality gate to the AI-agent evaluation framework.
Where is the e2e reference for a voice pipeline's model and budget chain?
The STT/LLM/TTS time-and-cost budget walks the per-step cost chain a model swap touches, and the LLM spend governor covers the spend-control surface.
The takeaway
The OpenAI deprecation cycle is a sorted calendar, not a blanket. Sort "surface sunset" (Assistants, hosted Evals, Agent Builder) from "model retirement" (legacy snapshots) from "policy restriction" (fine-tuning windows), mark Chat Completions as the primitive that stays, then re-author, re-bump, or monitor accordingly. The LLM choke-point in an AI-voice console lives in exactly one place; make it swappable and the calendar is a schedule, never an emergency.