Zapier has published a detailed account of how ChatGPT models can be wired into its automation platform, describing two distinct integration paths rather than a single new product. The post, credited to Elena Alston with an original publication in May 2023 and a most recent update in September 2026, sets out the difference between adding an AI step to an existing Zap and letting ChatGPT drive automations directly from its own chat window.
The first path is AI by Zapier, which the post describes as a way to drop a ChatGPT step into a workflow wherever AI reasoning is useful, without the user needing their own OpenAI account. Zapier frames this as a deliberate design choice: deterministic automation handles routine routing, such as sending a form submission to a spreadsheet, while AI is reserved for tasks that need reasoning.
Zapier's own examples of those reasoning tasks include generating a follow-up email from a sales call transcript, extracting action items from a meeting note, and categorizing an inbound support request. The post walks through a support-ticket scenario in which new Airtable records trigger a ChatGPT step with a classification and prioritization prompt, after which high-priority tickets are filtered and routed to a Slack channel.
Control over the AI step is presented as configurable rather than fixed. According to the post, users can tune the prompt, attach knowledge sources, enable tool calling, and test output before relying on it. Model selection happens through a dropdown in the visual workflow editor, and Zapier states that new OpenAI models are added to the lineup as they are released.
The billing model is the most concrete operational detail in the post. Zapier says the cost of an AI step comes out of the user's Zapier task count rather than a separate OpenAI bill, so no tokens are consumed on the user's own account. It states that some AI steps use a single task like any other step, while advanced models use three or five tasks. Users who want a model not offered in AI by Zapier can bring their own key and pay their provider directly.
Zapier also positions model choice as reversible. The post says a user who decides an OpenAI model is not behaving as wanted can swap it for a model from OpenAI or another provider without rebuilding the workflow. It recommends AI by Zapier over the dedicated Zapier ChatGPT integration, which uses the user's own OpenAI account, citing prompt testing, knowledge sources, agentic tool calling, the absence of token costs, and easier model switching.
The second path is Zapier MCP, which the post describes as giving ChatGPT governed access to more than 9,000 apps so it can take actions directly from the chat window or from ChatGPT Work. Setup is described as connecting once, specifying which apps and actions ChatGPT may access, and then issuing natural-language instructions.
Zapier's illustrative prompts include pulling the last five HubSpot opportunities, calculating weighted pipeline and pushing a summary to Slack; catching up on missed email and adding flagged items to Todoist; checking Glean for trending company topics and comparing them against an Airtable editorial calendar; and drafting a follow-up email from a Salesforce call note plus creating a calendar invite. These are examples supplied by Zapier, not documented customer results.
Governance is the thread Zapier uses to tie the two paths together. The post says the connection runs through Zapier's managed authentication, so ChatGPT does not ask for an API key, credentials stay governed in one place, and actions run through the same permissions applied elsewhere in the stack. It adds that Zapier MCP works with ChatGPT and other agent harnesses that support MCP, so permissions do not have to be rebuilt if a team switches AI tools.
Zapier lists several automation use cases beyond support triage: enriching and scoring inbound leads before CRM entry, extracting project details from status messages into a tracking sheet, drafting social copy from a Slack idea, summarizing sales call transcripts with multi-step reasoning and logging structured fields to a CRM, running sentiment analysis or theme extraction on survey responses, and escalating incidents against a defined severity rubric.
For freelancers, designers and developers, the practical question is where the boundary sits between work that needs a model and work that does not. Zapier's own framing is that deterministic steps should carry routine routing and AI should fire only when a task requires reasoning, which matters for anyone estimating automation costs per client project. The task-based pricing described in the post is the detail most likely to affect scoping, because it replaces per-token spend with a countable unit that can be forecast.
The model-swapping claim is also relevant to client work. If a workflow can move between providers without being rebuilt, an agency or solo developer can respond to a client's model preference or a provider change without re-engineering the automation. That is Zapier's stated benefit, not an independently verified outcome, and the post does not quantify how much rework is actually avoided.
Several things remain unresolved in the supplied material. Zapier does not publish per-task pricing, plan tiers, or regional availability in the post, so the real cost of the three- and five-task advanced models cannot be calculated from the evidence. The post also does not state which specific OpenAI models are available at any given moment, only that new releases are added to the lineup.
Nor does the post offer measured results. The support-ticket, lead-scoring and transcript-summarization scenarios are described as ways a user might apply the tools, not as tested deployments with accuracy, time-saved or error-rate figures. Zapier references AutomationBench, its benchmark for evaluating AI models on business workflows, as a way to compare models, but the post does not report any benchmark results.
The MCP path carries its own open questions. Zapier says ChatGPT can act across 9,000+ apps once permissions are configured, but the post does not describe failure handling, audit logging detail, or what happens when an agent takes an unintended action. It also does not specify which MCP-compatible agent harnesses beyond ChatGPT have been validated.
The update date is itself a limitation for readers trying to judge currency. The post states it was originally published in May 2023 and most recently updated in September 2026, but the evidence does not break out which features existed at which point, so it is not possible to tell from this material alone whether AI by Zapier, Zapier MCP, or the task-based pricing described are new or long-standing.
Taken together, the post is best read as Zapier's own explanation of its ChatGPT integration options rather than an announcement of a dated release. The substantive, checkable claims are architectural and commercial: two integration paths, model selection by dropdown, task-based rather than token-based billing for AI by Zapier, bring-your-own-key as an alternative, and managed authentication for MCP actions.
For a freelancer deciding whether to build client automations on this stack, the sensible next step is to verify current model availability, task pricing and plan requirements directly with Zapier, since none of those are settled in the source. The post establishes the shape of the offering and the vendor's stated rationale; it does not establish cost, performance or reliability for any particular workflow.