OpenAI Model Deprecation, October 23: Budget the Migration, Not the Rate Card

On October 23, 2026, OpenAI switches off 11 API models, including gpt-4, gpt-4-turbo, o1, o3-mini and o4-mini, plus five families of fine-tuned models. On 5 of the 8 text migration paths OpenAI names, the replacement costs more per output token than the model it replaces. This OpenAI model deprecation is cheap to read about and expensive to absorb. The real cost sits in re-running evals, retuning prompts and re-checking outputs, and none of that is on a pricing page.

The short answer

If you call any of these models in production, you have 11 days. The replacement OpenAI names is gpt-5.6-sol for gpt-4, gpt-4-turbo, o1 and o3-mini, gpt-5.6-terra for o4-mini and gpt-3.5-turbo, and gpt-5.6-luna for gpt-4.1-nano. Budget engineering time for re-running evals, resetting reasoning effort and fixing tool calls, then check the new per-token bill.

What the OpenAI model deprecation on October 23 shuts down

OpenAI announced this batch on 22 April 2026, by email and on its deprecations page, under the heading "Legacy GPT model snapshots". The stated reason was to improve reliability and make model choice simpler. The notice gave each model a substitute, and that substitute column is where your budgeting should start.

OpenAI's policy promises at least 6 months of notice for generally available models, and this batch got 6 months to the day. That is consistent with the policy. It is also the minimum, which matters if your release cycle runs quarterly and you only noticed in September.

Prices are OpenAI's standard tier, read from its pricing page on 12 October 2026. GPT-5.6 rates apply to prompts of 272K input tokens or fewer, and the gpt-5.6-sol rate is labelled promotional, available at least through 21 November 2026. Cached input and batch discounts are excluded. ft-o4-mini follows o4-mini to gpt-5.6-terra.

The official list contains at least one error

Read the table on the deprecations page with care. gpt-4-1106-preview appears in the October 23 batch, but the same page also lists it in an earlier batch that shut down on 26 March 2026. A developer pointed this out on OpenAI's forum thread for the notice on 23 April, the day after it went out.

The same thread flagged that the original email ran two model names together and briefly included an embeddings model that was later withdrawn. None of this changes the date. It does mean you should inventory your own model strings rather than trust any single list, including this one.

Fine-tuned models retire with their base

Five fine-tuned families go on the same day: ft-gpt-3.5-turbo, ft-gpt-4, ft-gpt-4.1-nano, ft-babbage-002 and ft-davinci-002. OpenAI's rule is that inference on a fine-tune stays available only until its base model is deprecated. The recommended path is a new base model, not a migrated fine-tune, and the pricing page notes that OpenAI is winding down its fine-tuning platform. If your quality depended on a tune, you are rebuilding it through prompting or another provider. That is a project, not a config change.

What the rate card says about the gpt-4 shutdown

The headline story for most readers is good news. gpt-4 and o1 users move to a model that costs a third as much per output token, and gpt-4-turbo users save a third. If your traffic is concentrated on those three, the gpt-4 shutdown lowers your unit price before you change anything else.

Who saves, and who pays more

The pattern is that OpenAI has collapsed its catalogue into three tiers, and the cheapest old models have no equally cheap successor. The most striking case is o3-mini, which OpenAI maps to its flagship gpt-5.6-sol rather than to the mid tier. At list price, that path costs 3.6 times more per input token and 4.5 times more per output token.

My view is that most o3-mini users should not follow the official mapping at all. OpenAI's own upgrade guide tells teams coming from latency-sensitive flows to "evaluate Luna and Terra first; use Sol only if quality requires it". Test o3-mini traffic on Terra before you accept a substitution that more than quadruples your output bill.

Sol's price is a promotion with a date on it

The $4 input and $20 output rates for gpt-5.6-sol carry a note on OpenAI's pricing page: promotional pricing is "available at least through November 21, 2026". The page does not state what the rate becomes afterwards. Any business case you write for a Sol migration should therefore show a second scenario at a higher price. The cheaper number expires 29 days after your forced switch.

Per-token price is also only half of the cost equation. Your bill is price multiplied by tokens, and the next section is about why the token count is the half that moves. The broader version of that argument, that falling token prices and rising AI bills coexist, is laid out in the analysis of why cheap tokens still produce rising AI bills.

Where the LLM migration bill actually lands

An LLM migration has two kinds of cost. The visible one is the price per token. The invisible one is every assumption your code, prompts and tests made about the old model's behaviour, each of which now has to be checked by a person.

OpenAI's upgrade guide is unusually blunt about this. It says "do not perform a blind model-string replacement", and it says never to call a migration complete "merely because model strings changed". The guide's own recommended process runs to a five-arm validation matrix. That is a vendor telling you the work is real, and you should believe it.

Reasoning effort now defaults to medium

GPT-5.6 supports six reasoning effort levels, from none to max, and if you omit the setting it defaults to medium. The guide warns that many older mini and nano integrations effectively ran with no reasoning at all. Swap the model string without setting effort, and a request that used to answer instantly now thinks first.

That has two costs. Latency rises, which can break timeouts set for the old model. Token use rises too, because OpenAI's reasoning models bill their internal reasoning as output tokens. On an o4-mini path already priced at 2.73 times the old output rate, an unplanned jump in output volume compounds the increase. Set effort explicitly on every call before you measure anything.

Tool calls on Chat Completions can simply fail

This is the trap I would check first. The upgrade guide states that function tools in Chat Completions are compatible only with an effective reasoning effort of none. Because the default is medium, a Chat Completions call that passes tools and sets no effort is, in the guide's own word, unsafe.

The fix is small: send reasoning_effort set to none, or move the flow to the Responses API, which OpenAI recommends for reasoning, tool and multi-turn work. Moving endpoints is not small, though. Request and response shapes differ, field names differ, and every parser downstream has to be retested.

Prompts need trimming, not porting

Prompts written for gpt-4 tend to be long, defensive and repetitive. OpenAI's guidance for GPT-5.6 runs the other way: state each instruction once, drop generic lines such as "think step by step", and expose only relevant tools. The guide reports that leaner prompts scored roughly 10 to 15% better in OpenAI's internal evals while using 41 to 66% fewer tokens.

Treat those figures as directional. They are OpenAI's own internal results, with no sample or task list published. The direction is still useful: a ported prompt is probably both worse and more expensive than a rewritten one. The guide also changes prompt caching. It adds a cache-write charge of 1.25 times the uncached input rate, so a prompt layout tuned for the old cache can lose hits on the new one.

Your LLM evaluation harness may be retiring too

Here is the scheduling problem that most migration write-ups miss. On 3 June 2026, OpenAI announced the deprecation of its own Evals platform. Existing evals become read-only on 31 October 2026, and the dashboard and API shut down on 30 November 2026. Graders built for eval workflows are part of the same transition.

So a team that uses OpenAI Evals to compare old and new models has a narrow window. The old model disappears on October 23. The harness that would prove the new model is equivalent freezes 8 days later and disappears 38 days later. OpenAI's evals guide points users toward Promptfoo as a migration path, which is a second tooling project landing on the same team.

The larger point is that a baseline only exists while the old model does. Capture old-model outputs on a representative sample now, store them outside any vendor platform, and you can still score the new model against them in December. How to build that kind of suite from scratch is covered in the guide to building an LLM eval suite that survives a model swap.

The LLM migration cost checklist

This checklist is built from OpenAI's deprecation notice, its GPT-5.6 upgrade guide and its pricing page. Each row is a line item that costs time, money or both. The owner column is a suggestion for a small team, not a rule.

Two rows deserve emphasis. The baseline row is the only one with a hard deadline that cannot be recovered after it passes. The tier row is the only one that can turn a cost increase into a cost decrease. That makes it the best return on an hour of testing in the list.

If you run many models across providers, the logging work in the tool-call and caching rows is easier with tracing already in place. The trade-offs between the main options are compared in the comparison of LangSmith, Braintrust and Arize for LLM observability.

Where this argument is weakest

This post argues that migration labour, not token price, dominates the cost of October 23. There are good reasons to think that is wrong for some readers, and they should be stated plainly.

For gpt-4 and o1 shops, the savings may dwarf the work

A team spending heavily on gpt-4-0613 at $60 per million output tokens moves to $20 on Sol. If that spend is large, a few weeks of engineering time pays for itself quickly, and the rate card really is the headline. My argument is strongest for teams on mini and nano models, where both the labour and the price move against you. It is weakest for teams still on the expensive 2023 generation.

Drift happens without any deprecation

A fair objection is that model behaviour changes anyway. A peer-reviewed study in the Harvard Data Science Review compared two versions of the same named model. Lingjiao Chen, Matei Zaharia and James Zou found that GPT-4's accuracy on identifying prime numbers fell from 84% in March 2023 to 51% in June 2023. That task is narrow and the data is from 2023, so treat it as evidence that drift exists, not as a measure of how large it is today.

If drift is constant, then a forced migration is just a scheduled version of work you should be doing anyway. I accept that. It strengthens the case for a stored baseline and a harness you own, and it weakens the case that October 23 is uniquely costly.

The last limit is simpler. Nothing here measures your traffic. The price multiples are per token, and your bill depends on token counts that will change after migration in ways only your own logs can show.

How to budget the LLM migration in hours, not tokens

A migration budget has three parts, and only one of them is on a pricing page. Write each down separately so that nobody quietly assumes the other two are zero.

The first part is labour: the checklist rows above, multiplied by your team's loaded hourly cost. The second is the new run rate: last month's tokens by model, repriced at the new tier, with a second scenario where output tokens rise because of reasoning. The third is risk: the cost of a regression reaching customers before you catch it, which is the part an eval suite exists to shrink.

I would also budget for the next one. OpenAI's deprecations page already lists gpt-5, gpt-5-mini, gpt-5-nano, o3 and o3-pro for shutdown on 11 December 2026. A team that migrates o3-mini to o3 to avoid Sol's price will be migrating again in seven weeks. The general case for treating switching cost as a recurring line item rather than a one-off is made in the breakdown of what switching AI models really costs.

One more option is worth naming, because it is in OpenAI's own policy. The deprecations page says some customers may be able to provision dedicated capacity to keep access after a shutdown date by contacting sales. For a regulated workflow that cannot be revalidated in 11 days, that call is cheaper than a rushed migration. I would make it this week.

Frequently asked questions

Where to start before October 23

Today, search every repository and config store for the 11 retiring model names and their aliases, and list each call site with its monthly token volume. That list is your budget, your test plan and your risk register in one document.

Then, before the 23rd, run 200 to 500 representative requests through each old model you still use and save the outputs to storage you own. Everything else on the checklist can happen in November. That step cannot.

Related reading