OpenAI AgentKit
Eight months from launch to deprecation notice. Agent Builder and the Evals platform both shut down on 30/11/2026, the export is lossy by OpenAI's own admission, so anything you start here you build twice.
PRICING
| Agent Builder, Agents SDK, Guardrails | Free | no seat licence and no enterprise gate; you pay for the tokens and tool calls the workflow consumes |
| ChatKit upload storage | $0.10/GB-day | after 1 GB free per account per month; started being charged 01/11/2025, a month after a launch that promised 'included with standard API model pricing' |
| Connector Registry | Gated | requires a Global Admin Console tenant, itself limited to eligible ChatGPT Business, Enterprise or Edu organisations |
| Workspace Agents (the migration target) | Credit-based | requires a ChatGPT Business, Enterprise or Edu workspace; moved to credit pricing 06/07/2026, rates not verified |
The headline is honest — there is no AgentKit licence — but the two things that cost you are not on the price list: the migration off Agent Builder before 30/11/2026, and self-hosting a ChatKit backend once the hosted one disappears. Developers on OpenAI's own forum benchmarked replacements at 2.5-3.4s cold starts on Lambda and about 20s on a plain VM, pushing them toward an always-on box at €15-30 a month.
checked 2026-08-26 · vendor pricing page
Element scores
Strengths
Taken purely as engineering, the parts are good. Tracing is on by default and captures model calls, tool calls, guardrails and handoffs in one run view. MCP support is deep and genuinely open: an MCP node on the canvas, hosted and SDK-managed servers over stdio or streamable HTTP, per-tool approval items, a secure tunnel for private servers, and eight first-party connectors. Hosted vector stores and the file search node make retrieval a two-minute job. The cost ladder is real, roughly 20x between gpt-5.6-luna and gpt-5.6-sol, and the SDK lets you mix model sizes per agent. There is no licence fee: you pay tokens. And the Agents SDK, which is where OpenAI is now pointing everyone, got substantially better in April 2026 with native sandboxing across seven providers, durable execution, subagents and configurable memory — it is provider-agnostic and will talk to 100+ non-OpenAI models, which the visual builder never would.
Honest dings
On 03/06/2026 OpenAI deprecated Agent Builder, the Evals platform and reusable prompt objects on the same day. All three shut down on 30/11/2026, Evals goes read-only on 31/10/2026, and in the docs navigation Agent Builder and Evals now sit under a heading labelled 'Legacy APIs'. That is eight months from DevDay launch to deprecation notice, and it is the second major OpenAI agent deprecation in a year, after the Assistants API. Three consequences to price in. The export is lossy by OpenAI's own wording: it 'does not convert your workflow graph or guarantee that every behavior transfers unchanged', and workflows with strong determinism 'may not migrate faithfully'. The Evals exit is worse, because OpenAI Support confirmed not every eval type exports as a runnable Promptfoo config. And the OpenAI-hosted ChatKit backend has no replacement at all, so the recommended integration path becomes self-hosting. One paying developer running Agent Builder and ChatKit for multiple clients called the migration 'economically unviable' on OpenAI's own forum. Beyond the shutdown: the canvas is OpenAI-models-only, there is no audio or voice node anywhere in it, and Zenity Labs' security review found the guardrails classify rather than correct — they label the fire without putting it out.
Sources (23) — every claim traceable
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- developers.openai.com/api/docs/deprecations — The load-bearing official source: Agent Builder, the Evals platform and reusable prompt objects all deprecated 03/06/2026 with a 30/11/2026 shutdown, Evals read-only from 31/10/2026, 'ChatKit remains available', and the 07/05/2026 self-serve fine-tuning wind-down (accessed 2026-08-26)
- openai.com/index/introducing-agentkit — Official launch post (06/10/2025), now carrying a wind-down banner: verbatim tagline, the original five-component list, 'All of these tools are included with standard API model pricing', the never-shipped standalone Workflows API, and every customer figure quoted in this audit (Ramp 70%, LY under two hours, Canva two weeks, Carlyle 50%/30%) (accessed 2026-08-26)
- developers.openai.com/api/docs/guides/agent-builder — Official Agent Builder guide: deprecation banner, Legacy APIs placement, workflow autosave and publish-to-version model, the Evaluate button driving trace graders, and the Code export path (accessed 2026-08-26)
- developers.openai.com/api/docs/guides/node-reference — Complete official node inventory (Start, Agent, Note, File search, Guardrails, MCP, If/else, While, Human approval, Transform, Set state) with CEL expressions and typed edges. Primary evidence for the documented absences behind the Mm, Cx and Fc scores: no audio or voice node, no memory node, no parallel fan-out node (accessed 2026-08-26)
- developers.openai.com/api/docs/guides/agent-builder… — Official migration guide and the source of the lossy-export admission: the export 'does not convert your workflow graph or guarantee that every behavior transfers unchanged', and workflows with strong determinism 'may not migrate faithfully' to a workspace agent (accessed 2026-08-26)
- developers.openai.com/api/docs/guides/agent-builder… — OpenAI's own risk admissions behind the Gr and Sm scores: guardrail nodes 'are not foolproof', agents 'can still make mistakes or be tricked', structured outputs mandated as a security control, tool approvals to be left on for MCP, and GPT-5 or GPT-5-mini recommended for higher-risk workflows because smaller models are weaker against prompt injection (accessed 2026-08-26)
- developers.openai.com/api/docs/guides/agents/integr… — Official evidence for the Tr score: tracing on by default, what a default trace contains (model calls, tool calls, handoffs, guardrails, custom spans), hosted versus SDK-managed MCP servers — and the absence of any documented OpenTelemetry exporter (accessed 2026-08-26)
- developers.openai.com/api/docs/guides/agent-evals — Official evaluation decision page: traces as the entry point for SDK-based apps, and the statement that Agent Builder traces remain visible only 'during the transition window' (accessed 2026-08-26)
- developers.openai.com/api/docs/guides/trace-grading — Official: what trace grading scores (decisions, tool calls, reasoning steps) and the Grade-all flow routing into the evaluation dashboard that shuts down on 30/11/2026 (accessed 2026-08-26)
- developers.openai.com/cookbook/examples/evaluation/… — Official cookbook confirming the replacement path: 'OpenAI is winding down the Evals product and recommends Promptfoo for continuing and extending your evaluation workflows' (accessed 2026-08-26)
- developers.openai.com/api/docs/guides/agents/models — Official: reasoning-effort control via model settings, mixing model sizes per agent, non-OpenAI models routed through the SDK provider/adapter surface, and the explicit separation of the voice-session path from agent workflows (accessed 2026-08-26)
- developers.openai.com/api/docs/guides/retrieval — Official vector store documentation behind the Vx score: CRUD, 500-file batches, attribute filtering, 800-token default chunks with 400 overlap, hybrid search with tunable embedding and text weights, score thresholds, expiration policies and per-store rate limits (accessed 2026-08-26)
- developers.openai.com/api/docs/guides/tools-connect… — Official evidence for the Pc score: the eight first-party connector IDs, the note that OpenAI prioritised services lacking official remote MCP servers, the per-request OAuth model, and the MCP approval request and response items (accessed 2026-08-26)
- developers.openai.com/api/docs/pricing — Live pricing verification: the single AgentKit line item (ChatKit file and image upload storage, $0.10/GB-day after 1 GB free), gpt-5.6 sol/terra/luna token rates and cached-input discounts, file search $2.50/1k calls, web search $10/1k calls, and 'OpenAI is winding down the fine-tuning platform' (accessed 2026-08-26)
- developers.openai.com/api/docs/guides/model-optimiz… — Official fine-tuning matrix behind the Ft score: SFT, vision and DPO on 2024-25 base models, RFT restricted to o4-mini-2025-04-16, and no GPT-5.x fine-tuning target listed (accessed 2026-08-26)
- developers.openai.com/api/docs/guides/custom-chatkit — Official self-hosted ChatKit path that survives the shutdown: implementing ChatKitServer.respond(), the Store and FileStore contracts you own, thread metadata as server-side state, widgets, actions and the documented client event set (accessed 2026-08-26)
- openai.com/index/the-next-evolution-of-the-agents-s… — Official April 2026 Agents SDK update, the strength behind the Fw score: native sandboxing across seven providers, the Manifest workspace abstraction, durable execution via snapshotting, subagents and configurable memory, plus OpenAI's own framing that model-agnostic frameworks trade flexibility against frontier-model capability (accessed 2026-08-26)
- github.com/openai/openai-guardrails-python — Official repo evidencing the Gr maturity flag: the header still reads 'OpenAI Guardrails: Python (Preview)' ten months after launch, with the Presidio third-party dependency disclosed (accessed 2026-08-26)
- labs.zenity.io/post/analyzing-the-security-risks-of… — Independent security teardown (Zenity Labs, 08/10/2025): excessive agency in a single node holding Drive, PayPal and custom MCP access; connector credentials generated by hand in an OAuth playground with nothing preventing an over-privileged token; and the core finding that the guardrails 'classify, not correct' — they label the fire but do not put it out (accessed 2026-08-26)
- community.openai.com/t/deprecation-notice-agent-bui… — Independent practitioner evidence: a developer running Agent Builder and ChatKit for multiple clients calling self-managed migration 'economically unviable', the warning not to move prompts into Prompt Objects because those die on the same date, and self-run cold-start benchmarks of 2.5-3.4s on Lambda versus about 20s on a plain VM (accessed 2026-08-26)
- community.openai.com/t/deprecation-notice-evals-wil… — Independent evidence that the Evals exit path is incomplete: a developer following the official Promptfoo cookbook on 02/07/2026 could not find the promised export, and OpenAI Support replied that 'not every evaluation type currently supports exporting as a Promptfoo configuration' (accessed 2026-08-26)
- mcp.directory/blog/openai-agentkit-deprecation-2026 — Independent analysis (22/06/2026) scoping exactly what is and is not deprecated, with the timeline and the structural conclusion: 'A hosted visual builder is a proprietary surface a vendor can sunset on its own schedule' (accessed 2026-08-26)
- venturebeat.com/orchestration/openai-unveils-agentk… — Independent launch coverage confirming the component list and availability at DevDay 2025, used as the baseline against which the June 2026 deprecation is measured (accessed 2026-08-26)