OpenAI GPT-4o retirement for Salesforce teams: the call

The call

The notice is on OpenAI’s deprecations page, dated 14 October 2026, with a shutdown date of 14 January 2027 and gpt-6-luna named as the replacement for all four models (vendor, OpenAI deprecations). Coverage followed the same day (independent, CellCog; Shelly Palmer).

What it is, and what it is not

What it is. An API shutdown of five snapshots: gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14, and gpt-4.1-mini-2025-04-14. OpenAI’s words on fine-tunes: they “follow the same shutdown schedule as their underlying base models” (vendor). OpenAI says some customers may get dedicated capacity after the date through sales (vendor, via CellCog).

What it is not. It is not a ChatGPT change; the notice covers the API only. It is not six months’ notice. CellCog reads OpenAI’s policy as at least six months for generally available models, with exceptions for safety or compliance, and OpenAI gave no reason here (independent, CellCog). And it is not a like-for-like swap. Luna is a reasoning model with effort settings from none to max, default medium, and reasoning tokens bill at the output rate (independent, ComputingForGeeks).

The number nobody is quoting

Coverage says Luna is cheaper. Per token, it is: $0.10 input and $0.50 output per million, against $2.50 and $10.00 for GPT-4o, $0.15 and $0.60 for GPT-4o mini, $2.00 and $8.00 for GPT-4.1, and $0.40 and $1.60 for GPT-4.1 mini (vendor list prices, via CellCog). The question is how many output tokens Luna can spend before the call costs more than the one it replaces.

Take a typical service call: 1,000 tokens in, 300 out, such as a case summary. Luna breaks even when its output reaches (old input price × 1,000 + old output price × 300 − $0.10 × 1,000) ÷ $0.50.

Luna on the same 1,000 in and 300 out, reasoning off, costs $0.25 per 1,000 calls (our maths). Moving off GPT-4o or GPT-4.1 is a large saving at almost any effort level. Moving off GPT-4o mini, the cheapest of the four, the saving is $0.08 per 1,000 calls and disappears at 160 reasoning tokens per call. Every 100 reasoning tokens adds $50 per million calls (our maths). Leave effort on the default and the cheaper model can be the dearer one.

The calendar is tighter than the price. From 15 October to 13 January is 65 weekdays; take out Thanksgiving, Christmas Day, and New Year’s Day and 62 US working days remain, before any year-end change freeze (our maths).

Missing

What OpenAI has not told an enterprise buyer, as of 15 October:

• A fine-tuning path. One independent spec sheet says Luna does not support fine-tuning (ComputingForGeeks); OpenAI’s notice is silent.
• A reason for three months instead of six.
• Any published comparison of Luna with GPT-4o or GPT-4.1 on the same tasks.
• Typical reasoning tokens per call at each effort level.
• Price, minimums, and term for dedicated capacity after 14 January.
• Tool calling with reasoning on. ComputingForGeeks reports Chat Completions tool calls work only at effort none.
• Dates for the same models on Azure OpenAI.

Who owns what

Retiring the model is OpenAI’s call. The callout that names it is ours.

In a Salesforce org the model name usually sits in an Apex string, a Named Credential’s request body, a Flow’s invocable action, or a MuleSoft config. The Named Credential and host stay the same; the model string, the effort setting, and the response parsing change. OpenAI’s own advice is to test representative workloads, check output formats and tool use, and roll out gradually (vendor, via Shelly Palmer). Apex that parses JSON from the reply is where a quiet format change becomes a failed case update.

Fine-tunes are a different problem. If Luna cannot be fine-tuned, the choice is prompt engineering on Luna, dedicated capacity at an unpublished price, or another vendor. That is a commercial decision, not a code change.

Monday morning

1. Step 1. Inventory. Search Apex, Named Credentials, Custom Metadata, Flows, and middleware for gpt-4o, gpt-4.1, and ft:, and match against OpenAI usage by model. Owner: Salesforce platform owner.
2. Step 2. Split out fine-tunes. List each one, what it does, and its volume; ask your OpenAI account team in writing for the Luna fine-tuning position and the dedicated-capacity price. Owner: AI lead and procurement.
3. Step 3. Shadow-test Luna. 200 real calls per use case at effort none and low; compare answers, JSON validity, tool calls, and output tokens per call. Owner: integration lead.
4. Step 4. Move the model name into Custom Metadata. Set effort explicitly and log output tokens per call, so the switch and the rollback are configuration, not a deploy. Owner: Salesforce developer.
5. Step 5. Cut over gradually, before any December freeze. Start with the lowest-risk flow and keep the old model as fallback until it disappears. Owner: product owner for each integration.

We move this to WATCH for fine-tuned workloads if OpenAI confirms no Luna fine-tuning and prices dedicated capacity. We drop the urgency if OpenAI moves the date.