OpenAI deprecated the last GPT-4o and GPT-4.1 snapshots on 14 October 2026. They stop answering on 14 January 2027 — 88 days from today — and fine-tuned models built on them stop with them. OpenAI names one replacement for all four: gpt-6-luna. That recommendation is right for GPT-4o mini and GPT-4.1 mini, where it is a price cut. It is a tier demotion for GPT-4o and GPT-4.1, whose like-for-like successor is GPT-6.1 Sol at $2/$10 — 25% more per output token than GPT-4.1’s $8, offset by a five-times-cheaper prompt cache. Whether that nets out cheaper or dearer depends on one ratio, below. And one capability has no OpenAI successor at any price: GPT-4.1’s flat-rate million-token context.
The six IDs to grep for, and their two different dates
Two separate deprecation entries are in play, which is why a single search misses half the work.
| Model ID | Shuts down | Days left | OpenAI’s named replacement |
|---|---|---|---|
gpt-4o-2024-05-13 | 23 Oct 2026 | 5 | gpt-5.6-sol |
gpt-4.1-nano-2025-04-14 | 23 Oct 2026 | 5 | gpt-5.6-luna |
gpt-4o-2024-08-06 | 14 Jan 2027 | 88 | gpt-6-luna |
gpt-4o-2024-11-20 | 14 Jan 2027 | 88 | gpt-6-luna |
gpt-4o-mini-2024-07-18 | 14 Jan 2027 | 88 | gpt-6-luna |
gpt-4.1-2025-04-14 | 14 Jan 2027 | 88 | gpt-6-luna |
gpt-4.1-mini-2025-04-14 | 14 Jan 2027 | 88 | gpt-6-luna |
Dates from OpenAI’s deprecations page, checked 18 October 2026. The 23 October models are part of a much larger batch — eleven models and five fine-tune bases going on one day — covered in full in OpenAI’s 23 October shutdown.
Three things that search misses:
- The floating aliases. OpenAI’s table names snapshots only. If your code says
gpt-4o,gpt-4o-mini,gpt-4.1orgpt-4.1-miniwith no date, the page does not say where that alias resolves after 14 January. Treat it as retiring, and pin an explicit ID so that you choose the replacement rather than inherit one. - Model names in data, not code. Routing tables, prompt templates, per-customer config rows, environment variables and saved evaluation runs all hold model strings that a code search will not reach.
- The programmatic check. OpenAI’s models endpoint now returns a
shutdown_datefield —YYYY-MM-DD, or null if nothing is announced. One scheduled job that lists your models and alerts on any non-null value replaces the practice of reading deprecation pages.
Every date here is also in our AI model retirement calendar, which publishes an .ics feed you can subscribe to.
What the replacement actually costs
Prices are per million tokens, from OpenAI’s pricing page, verified 18 October 2026.
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-4o | $2.50 | $1.25 | $10.00 |
| GPT-4o mini | $0.15 | $0.075 | $0.60 |
| GPT-4.1 | $2.00 | $0.50 | $8.00 |
| GPT-4.1 mini | $0.40 | $0.10 | $1.60 |
| GPT-6 Luna (named replacement) | $0.10 | $0.01 | $0.50 |
| GPT-6 Sol | $2.00 | $0.20 | $10.00 |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 |
Priced on one reference workload — a retrieval-backed extraction service running 10,000 requests a day, 8,000 input tokens and 600 output tokens each, half the input served from the prompt cache, over a 30-day month:
| Running today | Monthly bill | On GPT-6 Luna | On GPT-6.1 Sol |
|---|---|---|---|
| GPT-4o | $6,300 | $222 (−96%) | $4,320 (−31%) |
| GPT-4.1 | $4,440 | $222 (−95%) | $4,320 (−3%) |
| GPT-4.1 mini | $888 | $222 (−75%) | — |
| GPT-4o mini | $378 | $222 (−41%) | — |
Run your own numbers in the API cost calculator, which has this table’s models preloaded and flags the retiring ones.
For the two mini models, stop here: GPT-6 Luna is cheaper and newer, the recommendation is sound, and the only work left is an evaluation run. The rest of this page is about the other two.
The rule that decides whether GPT-6 Sol costs you more
GPT-4.1 and GPT-6 Sol both list at $2 per million input tokens. Everything else about the move pulls in opposite directions:
- Output costs more: $10 against $8, a 25% increase.
- Cached input costs much less: $0.20 on GPT-6 Sol and $0.10 on GPT-6.1 Sol, against GPT-4.1’s $0.50. GPT-4.1 discounts cached tokens by 75%; GPT-6.1 Sol discounts them by 95%.
So the migration is a price cut for prompt-heavy, cache-heavy work and a price rise for answer-heavy work, and the crossover is a single ratio. Write O for output tokens per request, I for input tokens, and C for the share of input served from cache. Setting the two effects equal:
GPT-6.1 Sol is cheaper than GPT-4.1 when C > 5 × (O ÷ I). For GPT-6 Sol, the multiplier is 6.7 instead of 5.
Which looks like this:
| Your output:input ratio | Shape of workload | Cache hit rate needed to break even |
|---|---|---|
| 1:40 | Document QA, long context, short answers | 13% |
| 1:20 | Retrieval with citations | 25% |
| 1:13 | The workload priced above | 38% |
| 1:10 | Classification with explanations | 50% |
| 1:7 | Summarisation | 71% |
| 1:5 or worse | Chat, drafting, code generation | Impossible |
The last row is the one worth internalising. Once your output exceeds 20% of your input, no cache hit rate makes GPT-6.1 Sol cheaper than GPT-4.1 — the output surcharge is larger than the entire cached-input saving even at a perfect cache. For GPT-6 Sol the threshold is 15%. Conversational products and code assistants sit well past it; retrieval pipelines sit well short of it. Two teams migrating off the same model will correctly reach opposite conclusions.
Note also that GPT-6.1 Sol and GPT-6 Sol differ only in the cached-input rate — $0.10 against $0.20, identical on everything else. There is no workload on which GPT-6 Sol is the better buy, which is a recurring feature of OpenAI’s rate card rather than an accident; we covered the pattern when GPT-6.1 Sol launched at the same headline $2.
The capability with no OpenAI successor
GPT-4.1 accepts 1,047,576 input tokens and bills all of them at $2 per million. No surcharge, no tier, no threshold. Its replacement does not do that, and neither does anything else OpenAI sells.
Every GPT-6 and GPT-5.6 model splits its rate card at 272,000 input tokens. Above that line OpenAI charges double the input rate and 1.5 times the output rate for the whole request — GPT-6 Sol goes from $2/$10 to $4/$15, GPT-6 Luna from $0.10/$0.50 to $0.20/$0.75. The context window is nominally larger at 1.05M tokens, but the top three quarters of it cost twice as much as GPT-4.1 charged for the same tokens.
On a 400,000-token prompt with a 2,000-token answer, uncached:
| Model | Cost per call | Window |
|---|---|---|
| GPT-4.1 (retiring) | $0.82 | 1,047,576, flat rate |
| GPT-6 Sol | $1.63 | 1,050,000, surcharge above 272K |
| Claude Sonnet 5.5 | $0.82 | 1,000,000, flat rate |
| Gemini 3.8 Flash | $0.31 | 1,048,576, flat rate |
Claude Sonnet 5.5 reproduces GPT-4.1’s long-context economics almost exactly, at the same $2/$10 with no threshold — see the Claude review and Claude vs ChatGPT for the non-price differences. Gemini 3.8 Flash is a third of the cost, with one dated catch: every current Flash rate doubles on 1 January 2027, taking that call to $0.62. One caveat on all cross-vendor arithmetic: the three labs use different tokenizers, so the same text is not the same token count. Re-measure on your own corpus rather than carrying GPT-4.1’s token counts across — the full treatment is in what long prompts actually cost.
If your prompts never approach 272,000 tokens, none of this applies and you can ignore it entirely.
Fine-tunes, and the deadline eight days before the deadline
OpenAI states that “fine-tuned versions of these models follow the same shutdown schedule as their underlying base models.” So a fine-tune on gpt-4o-2024-08-06 or gpt-4.1-mini-2025-04-14 stops on 14 January 2027 along with its base.
The date that actually binds is earlier. OpenAI’s self-serve fine-tuning deprecation disables the creation of new fine-tuning jobs for existing customers on 6 January 2027 — eight days before the base models go. A team that plans to re-train on a GPT-6 base in the first week of January has to start the job before the 6th, and a failed run leaves no time for a second attempt.
Do the inventory this month:
- List every fine-tuned model your organisation owns and the base each was trained on.
- For each, decide re-train, replace with prompting plus few-shot examples, or retire.
- For anything re-trained, book the job well before 6 January 2027.
- Keep the training set. Whatever you do next needs it, and it is the asset that is genuinely hard to recreate.
The migration, in order
- Grep for all six IDs, plus the four undated aliases, across code, config, data and infrastructure-as-code. Start with
gpt-4o-2024-05-13andgpt-4.1-nano— those have five days. - Export usage by model from the OpenAI dashboard and reconcile it against what you found. The gap is your shadow usage.
- Measure O, I and C per workload — output tokens, input tokens and cache hit rate. Three numbers; they decide the whole migration and most teams do not have them to hand.
- Evaluate GPT-6 Luna first, on every workload. It is 20x cheaper than Sol on input. If it clears your quality bar the decision is made; if it fails, you have learned that cheaply.
- Pick Sol by the ratio, not by habit. Use the rule above; for anything past the 1:5 line, price the move honestly and plan for a bill increase rather than discovering one.
- Price the long-context calls separately if you have any over 272,000 tokens, and include Claude Sonnet 5.5 and Gemini 3.8 Flash in that comparison specifically.
- Ship before mid-December. A January cutover competes with the holidays, the fine-tuning deadline and everyone else’s migration.
Current plan prices for the consumer and team products sit in the pricing tracker; if your OpenAI exposure is through an editor rather than the API, migrating off Cursor’s OpenAI models covers that path, and the ChatGPT review covers the subscription side.
What changed
- 18 Oct 2026: Guide published. Prices verified against OpenAI’s pricing page; dates against its deprecations page — see the 92-day notice analysis.
Frequently asked questions
When do GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini stop working?
14 January 2027. OpenAI's deprecations page lists five snapshots under its '2026-10-14: GPT-4 series model deprecations' entry: 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. After that date requests to them fail. A sixth ID catches people out: gpt-4o-2024-05-13 was deprecated separately back in April and retires on 23 October 2026, so a service pinned to that snapshot has days, not months. The same 23 October batch takes gpt-4.1-nano.
Is gpt-6-luna really the right replacement for GPT-4.1?
Not on capability. OpenAI names gpt-6-luna for all four models, but Luna is the small tier of the GPT-6 family at $0.10 input / $0.50 output per million tokens, while GPT-4o and GPT-4.1 are mid-tier. For gpt-4o-mini ($0.15/$0.60) and gpt-4.1-mini ($0.40/$1.60) the recommendation is honest and it is a price cut. For GPT-4o and GPT-4.1 the like-for-like model is GPT-6 Sol or GPT-6.1 Sol at $2/$10, which is 25% more per output token than GPT-4.1's $8. Benchmark Luna first anyway — if it passes your evaluations the migration is a 90% saving — but do not assume it.
Will moving from GPT-4.1 to GPT-6.1 Sol raise my bill?
It depends entirely on your cache hit rate, because the output rate goes up ($8 to $10 per million) while the cached-input rate goes down ($0.50 to $0.10). The breakeven is a single ratio: GPT-6.1 Sol costs less than GPT-4.1 when your cached share of input tokens exceeds five times your output-to-input ratio. A retrieval workload with 8,000-token prompts and 600-token answers breaks even at a 38% cache hit rate. A chat workload whose answers exceed 20% of its prompt length can never break even, because no cache rate is high enough. Measure your own two numbers before you budget.
What happens to my fine-tuned GPT-4o and GPT-4.1 models?
They shut down with their base models on 14 January 2027 — OpenAI states that 'fine-tuned versions of these models follow the same shutdown schedule as their underlying base models.' The trap is the date before it: OpenAI's self-serve fine-tuning deprecation disables the creation of new fine-tuning jobs for existing customers on 6 January 2027, eight days earlier. So the last day you can start a replacement fine-tune on a model OpenAI still serves is before 6 January, not before 14 January. If a fine-tune is load-bearing, inventory it now.
What do I lose that no OpenAI model replaces?
Flat-rate long context. GPT-4.1 accepts 1,047,576 input tokens and bills every one of them at $2 per million with no surcharge. Every GPT-6 and GPT-5.6 model applies a long-context tier above 272,000 input tokens: twice the input rate and 1.5 times the output rate on the whole request. A 400,000-token prompt costs about $0.82 on GPT-4.1 and about $1.63 on GPT-6 Sol — the same job, double the money, with no OpenAI option that avoids it. The honest answers are cross-vendor: Claude Sonnet 5.5 at a flat $2/$10 to 1M tokens, or Gemini 3.8 Flash at $0.75/$3.75, though Flash rates double on 1 January 2027.
How much notice is 92 days, and is that normal?
It is below OpenAI's own published floor for this class of model. The deprecations page commits to 'at least 6 months' notice for generally available models, three months for specialised variants such as Codex and chat models, and as little as two weeks for previews. GPT-4o and GPT-4.1 are generally available models and received 92 days. The same family's gpt-4o-2024-05-13 snapshot got 184 days in April. OpenAI's policy permits a shorter period where safety or compliance requires it; the page gives no reason here. The practical lesson is to pin dated snapshots and poll the shutdown_date field the API now exposes on the models endpoint rather than trusting the notice period.