The Zeroic Journal

GPT-4o deprecation: move your app before January 14

What shuts down, what gpt-6-luna changes in your requests, what it costs, and a plan for fine-tuned models in the weeks left.

On this page 8 sections
  1. What the GPT-4o deprecation shuts down
  2. Find every call site before you switch
  3. gpt-6-luna reasons by default
  4. Build a 50-100 case regression set
  5. What gpt-6-luna costs against GPT-4o
  6. Fine-tuned gpt-4o-mini needs a new plan
  7. How to decide what to switch to
  8. How FormulaBot scaled to 1.5M+ users with validation

The GPT-4o deprecation takes gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini out of OpenAI’s API on January 14, 2027, and fine-tuned versions go with them. OpenAI’s named replacement is gpt-6-luna, which lists at $0.10 per million input tokens against $2.50 for gpt-4o. Switch with reasoning effort set to none, then test.

That’s 62 days from today, November 13. OpenAI announced the shutdown on October 14, 2026, which gave 92 days of notice. If your product went live in 2024 or 2025, gpt-4o-mini is probably hardcoded somewhere: an API route, a Supabase Edge Function, an n8n node. Changing the model name takes a minute. The parameters and behavior that change with it take the rest of the time.

What the GPT-4o deprecation shuts down

OpenAI’s deprecations page lists the snapshots, the date and the replacement:

ModelSnapshotsShutdownReplacement
GPT-4ogpt-4o-2024-08-06, gpt-4o-2024-11-20Jan 14, 2027gpt-6-luna
GPT-4o minigpt-4o-mini-2024-07-18Jan 14, 2027gpt-6-luna
GPT-4.1gpt-4.1-2025-04-14Jan 14, 2027gpt-6-luna
GPT-4.1 minigpt-4.1-mini-2025-04-14Jan 14, 2027gpt-6-luna

Three more details on the same page change the plan:

  • Fine-tunes follow their base model. OpenAI says fine-tuned versions “follow the same shutdown schedule as their underlying base models.” A fine-tuned gpt-4o-mini stops on January 14 too.
  • New fine-tuning jobs end first. From January 6, 2027, active existing customers can no longer create new fine-tuning jobs. The gpt-6-luna model page lists fine-tuning as not supported, so there is no self-serve way to fine-tune the replacement.
  • Some newer models go earlier. gpt-5, gpt-5-mini, gpt-5-nano, o3 and o3-pro shut down on December 11, 2026. If you moved part of your app to one of those, that date comes first.

Voice products have their own rows. The gpt-4o realtime and audio models shut down on January 20, 2027, and gpt-4o-transcribe and whisper-1 on February 26, 2027. If you run a voice agent, plan those alongside this one.

Find every call site before you switch

The model name is rarely in one place. Before changing anything, list every place your product calls OpenAI and which model each call names.

  • Your code. Search the repo for gpt-4o, gpt-4.1 and 4o-mini. Include serverless functions, background jobs and one-off scripts, not only the main API. If the app started in Lovable or Bolt, search its GitHub repo the same way, including any Supabase Edge Functions.
  • Config and the database. Model names often live in environment variables, a settings table or a per-customer config row. A code search won’t find those.
  • Automation tools. In n8n, each OpenAI Chat Model node has its own Model setting and options such as Sampling Temperature. Open every workflow that uses OpenAI and note both. Check any Zapier or Make step that calls OpenAI as well.
  • The API itself. The Models API returns a shutdown_date for each model, “or null if not announced.” A short script that checks every model name you found against that field turns the list into a test you can rerun.

Don’t count on the bare alias gpt-4o quietly pointing somewhere new. The deprecation entry names the dated snapshots and says nothing about a redirect. Plan as if every request naming one of these models fails on January 14.

gpt-6-luna reasons by default

The replacement changes how requests behave. The gpt-6-luna page lists reasoning effort levels of none, low, medium, high, xhigh and max, with medium as the default. GPT-4o and GPT-4o mini answer without a reasoning step. Swap only the model name and every request starts reasoning at medium effort. Three things follow.

Sampling parameters have to go. OpenAI’s GPT-6 guide says to drop sampling controls: “When reasoning effort is not none, remove temperature, top_p, and top_logprobs.” On Chat Completions, logprobs goes too. Most code written for gpt-4o-mini sets temperature.

Tool calls on Chat Completions need effort none. The same guide says GPT-6 Sol and GPT-6 Luna “support function calling in Chat Completions only with reasoning_effort: "none".” Most gpt-4o-mini apps with tools still use Chat Completions. Either keep it and set effort to none, or move that call to the Responses API if it needs reasoning and tools together.

Hidden tokens cost money and can empty a reply. OpenAI’s reasoning guide says reasoning tokens are billed as output and count against your output limit. When the limit is hit, the response comes back incomplete, which “might occur before any visible output tokens are produced.” An output cap of a few hundred tokens that suited gpt-4o-mini can return nothing. OpenAI suggests reserving at least 25,000 tokens while you experiment.

That’s why the first move is effort none:

from openai import OpenAI

client = OpenAI()

resp = client.chat.completions.create(
    model="gpt-6-luna",
    reasoning_effort="none",  # closest to gpt-4o-mini; keeps temperature and tools valid
    temperature=0.2,
    messages=messages,
    tools=tools,
)

At effort none, your request code keeps working and you compare like with like. Raise the effort later, one call type at a time, where your tests show it helps.

Build a 50-100 case regression set

Don’t move production traffic on the strength of a few manual tries. For each call type, pull 50-100 requests from your logs with the output the old model gave and a note on whether that output was right. Include the awkward ones: long inputs, other languages, cases where the old model refused or got it wrong.

Run the set through gpt-6-luna and compare five things:

  • JSON validity. If you parse model output, count parse failures. gpt-6-luna supports Structured Outputs, so if your JSON depends on prompt instructions alone, move to a strict schema as part of this change.
  • Length and tone. A newer model writes differently. If your UI has fixed-size cards or users expect a certain voice, read the outputs side by side.
  • Refusals. Count requests the new model declines that the old one answered, and the reverse.
  • Tool calls. Check each tool still fires with the right arguments, and that the model doesn’t start calling tools it used to skip.
  • Latency. Measure time to first token and total time at the effort level you plan to ship.

Keep the set in your repo and rerun it before every future model change. It’s the same method our guide to choosing GPT, Claude or Llama by task type uses to pick a model in the first place.

What gpt-6-luna costs against GPT-4o

On list price, the replacement is cheaper than everything it replaces. Per million tokens on standard processing, from OpenAI’s API pricing page:

ModelInputCached inputOutput
gpt-4o$2.50$1.25$10.00
gpt-4.1$2.00$0.50$8.00
gpt-4.1-mini$0.40$0.10$1.60
gpt-4o-mini$0.15$0.075$0.60
gpt-6-luna$0.10$0.01$0.50

Reasoning is what moves the bill. Take a call with 1,000 input tokens and 300 output tokens, run 100,000 times a month:

  • gpt-4o: $550 a month.
  • gpt-4o-mini: $33 a month.
  • gpt-6-luna at effort none: $25 a month.
  • gpt-6-luna with 1,000 reasoning tokens per call: $75 a month, more than double gpt-4o-mini.

A gpt-4o app saves heavily at any effort level. A gpt-4o-mini app saves only if most calls stay at none or low. Batch processing halves gpt-6-luna to $0.05 input and $0.25 output for work that can wait, and inputs over 272K tokens are billed at the higher long-context rate. Our post on cutting OpenAI API costs covers caching and batching in more depth.

Fine-tuned gpt-4o-mini needs a new plan

If your product runs on a fine-tuned gpt-4o-mini, January 14 removes it, and gpt-6-luna can’t be fine-tuned. For most teams the practical path is to move what the fine-tune learned into the prompt. Take 20-50 of your best training examples, put them in a fixed prefix as few-shot examples, and run the regression set.

Cached input on gpt-6-luna is $0.01 per million tokens. A long prefix stays cheap as long as it’s identical on every call and comes before anything that changes. If few-shot examples don’t close the gap, test a larger OpenAI model at effort none, or a provider that still offers fine-tuning. Make that call by mid-December so you have time to test the result.

How to decide what to switch to

Today you runSwitch toFirst setting
gpt-4o-mini or gpt-4.1-mini for classification, extraction or short repliesgpt-6-lunaEffort none
Any of the four on Chat Completions with function toolsgpt-6-lunaEffort none, or Responses API if it needs reasoning
gpt-4o or gpt-4.1 for multi-step work: agents, code, analysisgpt-6-lunaEffort low, then medium if tests miss
gpt-4o where gpt-6-luna fails the set at mediumgpt-6.1-terra ($1.00 input, $5.00 output)Responses API for tools
A fine-tuned gpt-4o-minigpt-6-lunaFew-shot examples in a cached prefix

A timeline that leaves margin:

  • By November 30: finish the call-site list and the regression sets. If you also use Agent Builder, that’s the same day it shuts down; our Agent Builder deprecation plan covers it.
  • December 1-18: run the sets, fix prompts and parameters, and settle any fine-tune replacement.
  • By January 7: move production traffic behind a flag you can flip back while the old model still answers.

How FormulaBot scaled to 1.5M+ users with validation

Agent products are where a model change bites hardest. For FormulaBot, now Better Analyst, our engineers designed its analysis agents with LangChain agents, LangGraph workflows and GPT-4o: an LLM-driven code execution engine that reasons through multi-step analysis, transforms the data and recovers when a step fails.

What makes a model change manageable is everything around the model. On FormulaBot we built the prompt engineering pipeline and the query validation layer, and split AI inference into its own independently deployable service as the platform grew to 1.5M+ users. With that setup, a model change is a change to one service, checked by a validation layer that already exists. That’s how we approach every OpenAI API integration and the rest of the work on our AI services hub.

If you want a second pair of eyes before January, an AI audit finds every model call in your product, ranks them by risk, and tells you which can switch at effort none and which need testing first.

Frequently asked questions

When does GPT-4o leave the OpenAI API?
GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down in the OpenAI API on January 14, 2027. OpenAI announced this on October 14, 2026. The dated snapshots listed are 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. Fine-tuned versions of these models stop on the same day.
What should I replace gpt-4o-mini with?
OpenAI names gpt-6-luna as the replacement for all four retiring models. It is a reasoning model with medium effort by default, so start with reasoning effort set to none. That keeps temperature and Chat Completions function calling valid and is the closest match to gpt-4o-mini. Raise the effort later only where your regression tests show it helps.
Will my fine-tuned GPT-4o mini model keep working?
No. OpenAI says fine-tuned versions follow the same shutdown schedule as their base models, so a fine-tuned gpt-4o-mini stops on January 14, 2027. gpt-6-luna does not support fine-tuning, and from January 6, 2027 existing customers can no longer create new fine-tuning jobs. Most teams move the fine-tune's behavior into few-shot examples in a cached prompt prefix.
Is gpt-6-luna cheaper than gpt-4o-mini?
On list price, yes. gpt-6-luna costs $0.10 per million input tokens and $0.50 per million output tokens, against $0.15 and $0.60 for gpt-4o-mini. Reasoning tokens are billed as output, though. At effort none the bill drops slightly. At higher effort, a call that adds 1,000 reasoning tokens can cost more than double what it did on gpt-4o-mini.
Why does gpt-6-luna return empty responses when gpt-4o-mini didn't?
Reasoning tokens count against your output limit. If a request hits the limit while the model is still reasoning, OpenAI returns an incomplete response, sometimes before any visible text. A small output limit that worked for gpt-4o-mini can leave nothing for the answer. Set reasoning effort to none, or raise the limit; OpenAI suggests reserving 25,000 tokens while you test.

References

  1. OpenAI - Deprecations
  2. OpenAI - GPT-6 Luna model
  3. OpenAI - Using GPT-6
  4. OpenAI - Reasoning models
  5. OpenAI - API pricing
  6. OpenAI API reference - Models
  7. n8n - OpenAI Chat Model node

More from the journal

Hear from Prashant within 24 hours.

Prashant Abbi Partner @ Zeroic

Still running gpt-4o-mini in production? Talk to a senior engineer who ships OpenAI features. In 30 minutes we'll map your call sites and tell you what the switch to gpt-6-luna involves.

Rather talk it through? Book a 30-minute call