November 2026 AI Deadlines: Agent Builder, Evals, Cursor, Sonnet 4.5

The short version

November 2026 is not a month of launches. It is a month of shutdowns, and several of them land on the same day.

On November 30, 2026, OpenAI shuts down three surfaces that teams quietly depend on: Agent Builder (the visual workflow canvas from AgentKit), the Evals platform, and reusable prompt objects including the v1/prompts API. Anthropic retires claude-sonnet-4-5-20250929 on the same date. And on November 12, OpenAI's contract supplying models to Cursor is scheduled to end.

Five different things change in one month: an agent-building surface, an evaluation platform, prompt storage, a model ID, and the model route inside a code editor. Handled one at a time, the month runs out. Handled as one dated checklist, it is a few days of work.

This piece orders them by date and separates what stops from what stays.

The deadline table

June 3, 2026 — OpenAI announces deprecation of Agent Builder, Evals, and reusable prompts.
October 31, 2026 — Existing evals become read-only. No new runs.
November 12, 2026 — Proposed shutoff of OpenAI models inside Cursor (OpenAI's own word: "proposed").
November 30, 2026 — Agent Builder shutdown, Evals dashboard and API shutdown, v1/prompts shutdown, Claude Sonnet 4.5 retirement.

What survives matters just as much. The Responses API, the models themselves, and ChatKit keep working. OpenAI's deprecations page says ChatKit remains available, and the Responses API is unaffected. Nothing about your prompts or your model calls is lost in principle — the hosted canvas and the hosted eval dashboard are what disappear.

The sharpest deadline already passed: Evals went read-only on October 31

This is the one teams miss. Evals did not wait for November 30. Since October 31, existing evaluations can be read but not run.

OpenAI points to Promptfoo, and there is a companion migration guide. The instinct is to export scores. Export the evidence instead:

fixture ID, so a run maps to a stable test case
baseline version, so you are not comparing against a moving target
candidate version, so behavior maps to a commit or branch
model and tool configuration
inputs and expected behavior
the run trace: tool calls, retries, decisions
score plus rubric notes, separating correctness, safety, cost, and style
cost and latency per run

You are not rebuilding a dashboard. You are preserving enough that someone else can replay the claim six months from now when the model behind it has changed twice.

Agent Builder: the export is not a conversion

OpenAI names two destinations: the Agents SDK (rebuild in code) and ChatGPT Workspace Agents (configure in natural language). Workspace Agents require an eligible Business, Enterprise, or Edu workspace plus permission to create agents, which rules them out for a lot of individual developers.

The trap is the export button. Agent Builder can emit Agents SDK code, but OpenAI's migration documentation is explicit that this does not convert the workflow graph and does not guarantee that every behavior transfers unchanged. "I exported it, so I migrated" is not a finished migration. It is a starting point that produces code you still have to review, test, and own.

A workable sequence:

Inventory. List every Agent Builder workflow, who uses it, how often, and what breaks if it stops.

Export the evals first. They have been read-only since October 31 — this is now data recovery.

Write each workflow down as text. Every node, prompt, tool, guardrail, branch condition, variable, and approval step. This document is the spec for the rebuild.

Sort by consequence. Business-critical goes to code. Simple internal assistants go to Workspace Agents. "When X happens in app A, do Y in app B" was never an agent — that is a workflow tool.

Run old and new side by side on real inputs from the last month before cutting over.

Add what the canvas never had: error alerting, retries that stop instead of looping, a human approval where mistakes are expensive, and logging.

One more check: ChatKit staying available does not mean a ChatKit app backed by an Agent Builder workflow keeps working. If your frontend points at a workflow ID hosted in Agent Builder, that backend goes away on November 30 even though the frontend product did not.

Cursor: November 12 ends bundled OpenAI models, not your editor

After SpaceX closed its acquisition of Anysphere, Cursor's parent, OpenAI notified SpaceX on August 28, 2026 that it would wind down the contract supplying models to Cursor, with a proposed shutoff of November 12, 2026. OpenAI says it is giving the maximum notice the contract allows and will not provide future models to Cursor during the wind-down.

Cursor CEO Michael Truell has said OpenAI models account for roughly 5% of Cursor traffic, which frames the blast radius correctly. For most users, the editor stays and one model option disappears.

If you explicitly pick OpenAI models: connect your own OpenAI API key (BYOK) or switch to Anthropic, Google, xAI, or DeepSeek. Your own key is not affected by the bundled contract.
If you live in Composer or another provider: check whether your default is still set to an OpenAI model and move on.
Your .cursorrules, prompts, and project context: owned by no vendor. A model swap is a config change, not a migration.

Claude Sonnet 4.5: swapping the model ID will fail in production

Sonnet 4.5 was deprecated on September 30, 2026 and retires on November 30, 2026 on the Claude API, Claude Platform on AWS, and Microsoft Foundry. Anthropic's policy promises at least 60 days' notice; this notice runs 61 days, one day above the floor.

The recommended replacement is claude-sonnet-5-5, released September 28, 2026, at $2 per million input and $10 per million output against $3 and $15 for Sonnet 4.5. Cheaper per token. And almost certainly not cheaper per request, because of what follows.

Sonnet 5.5 rejects several request shapes that 4.5 accepted. Each returns a 400:

Assistant prefill. 5.5 requires the conversation to end with a user message. Move format intent to structured outputs, and tone to the system prompt.
budget_tokens. Extended thinking is configured through thinking.type: adaptive and output_config.effort. Set effort explicitly rather than inheriting a default.
Non-default temperature, top_p, top_k. They hide in shared client wrappers and per-tenant config more often than in the handler you are reading.
Beta headers and parameters. Drop interleaved-thinking-2025-05-14; replace fine-grained-tool-streaming-2025-05-14 with eager_input_streaming: true per tool; replace output_format with output_config.format; on Bedrock, move computer use from computer_20250124 to computer_20251124.

Two silent changes do the financial damage. The tokenizer changed: the same text produces about 30% more tokens. And adaptive thinking is on by default, where a bare request on 4.6 and earlier ran without thinking — so that output now counts toward your tokens.

Images move more sharply. Sonnet 4.5 caps at 1568 pixels and 1,568 tokens; Sonnet 5.5 accepts up to 2576 pixels and 4,784 tokens. A 2000×1500 screenshot costs roughly 2.5 times the image tokens. If that detail is not worth it, downscale before sending.

And the date is platform-specific. Anthropic's dates cover the Claude API, Claude Platform on AWS, and Microsoft Foundry. Amazon Bedrock and Google Cloud set their own schedules. If you call 4.5 through Bedrock or Vertex, your deadline is not November 30 — and it is not on Anthropic's page either.

This week, in order

Find every model ID caller. Grep code, config, and infrastructure-as-code for claude-sonnet-4-5-20250929. Then open Claude Console → Usage → Export, filter the CSV by API key and model, and reconcile which keys still call it. The CSV tells you the key; grep tells you the repo.

Preserve eval evidence. Datasets, rubrics, and baseline runs — this is read-only recovery now, not a project.

Text-ify every Agent Builder workflow and sort them into code, Workspace Agents, and automation tools.

Check ChatKit backends for Agent Builder workflow IDs.

Check the Cursor default model. Decide BYOK or switch.

Write the rollback step before you cut over, and keep old and new running in parallel until it is boring.

What the month is actually saying

Line the three announcements up and they point one way: do not let production logic live only in a vendor's canvas or dashboard.

Agent Builder retires the canvas. Evals retires the dashboard. Reusable prompts retire the API. Every destination OpenAI names is code, a repository, or your own application. The Sonnet 4.5 retirement says the same thing about model IDs: they are not constants, they are dependencies with expiry dates.

The corollary is genuinely reassuring. If your prompts are in git, your evals replay from the repo, and your model IDs are centralized in config, November 30 is a one-line change. If they are not, the deadline is real work — and it is already November.