GPT-5 and o3 Leave OpenAI's API on Dec 11: Migration Guide

On December 11, 2026, OpenAI removes six model snapshots from its API: gpt-5-2025-08-07, gpt-5-mini-2025-08-07, gpt-5-nano-2025-08-07, gpt-5-pro-2025-10-06, o3-2025-04-16 and o3-pro-2025-06-10. The notice went out on June 11, 2026, which is the six months OpenAI promises for generally available models, to the day. Any request that names one of those IDs, or an alias that resolves to one, fails after that date.

The replacements OpenAI names are gpt-5.6-sol for GPT-5, o3 and both Pro variants, gpt-5.6-terra for GPT-5 mini and gpt-5.6-luna for GPT-5 nano. Two facts about those replacements are not on the deprecations page and matter more than anything that is. Each one with a published rate lists a higher per-token price than the model it replaces, by multiples of two to eight; the pro-mode rate for the two Pro rows is not published. And gpt-5.6-sol's current $4 and $20 per million is described on its own model page as promotional pricing that runs until November 30, 2026, eleven days before the old models go dark.

This post is the migration record for that date. The January 14, 2027 shutdown of GPT-4o and GPT-4.1 is a separate notice with its own replacement, and our January 14 migration guide covers it; the two cliffs are five weeks apart and most stacks have both.

The authoritative list is the entry dated June 11, 2026 on OpenAI's API deprecations page. It names six model families, gives one dated snapshot for each, and names one recommended replacement per row. Two of the rows carry a qualifier: the Pro variants map to gpt-5.6-sol with its reasoning mode set to pro, not to a separate Pro model.

The table below reproduces the mapping with the replacement's current list price beside it, because that is the column OpenAI's table does not have and the one a budget owner needs.

One line in the fine-tuning section of OpenAI's page deserves its own sentence: inference on fine-tuned models is disabled when the underlying base model is deprecated, so fine-tunes of these models stop with them. A fine-tune of gpt-5-mini built in September stops working on December 11 with everything else, and OpenAI's self-serve fine-tuning programme is itself winding down, with no new jobs for existing customers after January 6, 2027. A team that relied on a tuned GPT-5 variant has a harder migration than a team on the base model, because the replacement has to be re-tuned or prompted into the same behaviour.

The failure case here is the alias. A request to gpt-5 with no date suffix has always resolved to gpt-5-2025-08-07, because that is the only snapshot the page lists. No newer GPT-5 snapshot exists for the alias to roll forward to. The alias dies with the snapshot.

This notice does not move customers to a cheaper or equal-priced model, and the gap is large enough that it changes which replacement a cost-sensitive team should pick. The table puts each retiring model's list price beside the price of the model OpenAI recommends, with the multiple worked out.

Why the multiples run the wrong way is not stated anywhere by OpenAI, and we will not guess. What the pages do say is that the GPT-5.6 family was priced as a reduction against GPT-5.5, which the Sol page gives as 20% less on input and 33% less on output. GPT-5 was simply a cheaper generation. A team that chose gpt-5-mini for its $0.25 input price is being pointed at a model that costs eight times that on input.

One number in the table runs the other way and is worth knowing. Cached input on o3 is $0.50 per million; on gpt-5.6-sol it is $0.40. A reasoning workload with a long, stable system prompt and a high cache hit rate narrows the gap on the input side, which is the kind of arithmetic our prompt caching economics guide walks through. It does nothing for output, which is where o3's $8 becomes $20.

Ask OpenAI's documentation what should replace gpt-5 and you get three answers, depending on which page you open. All three are OpenAI's own, all three were current on November 5, 2026, and none of them is a measurement of how the model will do on your traffic.

Our view, and it is a process view rather than a model recommendation: the deprecations page mapping is the safe default because it is the one OpenAI will support if something breaks, and the two cheaper 6-series options are worth one eval run each before you commit a budget to the official route at two to eight times the old cost. Which of them wins on your workload is not knowable from a pricing table, and we have not run that test on anyone's traffic for this post. For the wider lineup and how the tiers relate, our GPT-6 Sol and Luna launch analysis carries the vendor-run benchmark figures, labelled as such.

Every retiring model is smaller than its replacement. The GPT-5 family lists a 400,000-token context window and 128,000 maximum output tokens, with GPT-5 Pro at 272,000 output. o3 and o3-pro list 200,000 context and 100,000 output. All three GPT-5.6 models list 1,050,000 context and 128,000 output, per their model pages.

The larger window carries a price rule that the old models did not have. The Sol page states that prompts over 272,000 input tokens are billed at twice the input rate and one and a half times the output rate for the whole request, not the part above the line. A pipeline that was comfortably inside gpt-5's 400K window with, say, 300K-token documents will cross the line on Sol and pay the uplift on every token. Our long-context price cliff analysis covers how the 200K and 272K thresholds differ across vendors.

The example that catches teams is the batch summariser. On gpt-5 it ran 350K-token inputs at $1.25 per million. On gpt-5.6-sol the same request is over the line and bills at $8 per million input, a 6.4× increase on a workload that looked like a 3.2× one in the price table. Knowledge cutoffs move too: GPT-5 lists September 30, 2024; o3 lists June 1, 2024; the 5.6 family lists February 16, 2026. A prompt that compensated for a 2024 cutoff by injecting recent facts can usually inject less.

A repository grep for the six snapshot strings is the first step and it is not enough, because most calls use the bare alias and several do not live in your code at all. The router below is the order we check in, from the place most teams forget to the one they remember.

The gateway row is where the single worst failure lives. A marketing team using an automation product that offers a "GPT-5" option has no snapshot in any file they control. Our LLM gateway architecture reference argues for owning that mapping yourself for exactly this reason: one routing table you control beats a dozen vendor defaults you cannot see.

The deprecations page now documents something that makes the manual calendar optional. OpenAI's list-models endpoint returns an optional shutdown_date field per model: a date string in YYYY-MM-DD form when a shutdown has been announced, or null when it has not. The same field is on the retrieve-model endpoint. A scheduled job that lists models and alerts on any non-null date inside your horizon replaces reading the page by hand.

The why is simple: the page is updated by OpenAI on its own schedule and read by you on yours, and the gap between the two is where surprises live. The failure case is trusting the field alone. It reports OpenAI's shutdown; it knows nothing about a gateway that maps its own names, or about a fine-tune whose base model is the one listed. Use the field for the models you call directly and the audit above for everything else.

Nothing in OpenAI's notice says the replacement behaves like the model it replaces, and this post makes no claim that it does. GPT-5 and o3 were trained on different data, with different reasoning defaults, more than a year before the 5.6 family. The only evidence that matters is a replay of your own recent requests through the candidate, scored the way you already score production. Our guide to testing a model on your own traffic is the procedure; the summary is one week of sampled requests, one scoring rubric that already exists, and a side-by-side run before any routing change.

Two things specific to this migration go on the checklist. First, output length: the 5.6 models are reasoning models with a 128,000-token output ceiling, and a prompt tuned to keep gpt-5 terse may produce longer answers, which at $20 per million output is the line item that moves. Second, structured output and tool schemas: confirm every JSON schema and function definition validates on the new model before December, because a schema rejection is a hard failure, not a quality drift. If the test turns up a regression you cannot prompt away, that is the point at which an AI transformation engagement earns its fee: the fix is usually in the surrounding pipeline, not the model.

December 11 is one of eight dated shutdowns OpenAI currently lists between now and April. A team doing the audit above once should do it against the whole list, because the second grep is nearly free and the second outage is not.

The October 23 shutdowns of o1, o3-mini, o4-mini and GPT-4 Turbo have already happened and are not in the table; a stack still calling those is already failing. The full history, including the July codex snapshot retirements, is in our model alias and retirement ledger.