Codex Snapshots Retire July 23: Your Migration Map

OpenAI’s Codex snapshot retirements land on July 23, 2026 — 13 model snapshots shut down in a single wave, including every Codex-lineage coding model before GPT-5.3: gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini, and gpt-5.2-codex. If any config, CI script, or agent backend still pins one of those IDs, it stops working in nine days.

To be clear about what this is: an already-announced schedule, not a surprise. OpenAI published the “Legacy GPT model snapshots” notice on April 22, 2026, giving roughly three months of runway — exactly what its own deprecation policy promises for specialized variants. The news value now isn’t the announcement; it’s that the deadline is close and the replacement mapping is genuinely not a find-and-replace.

This guide organizes OpenAI’s flat deprecation table into an actionable migration map: what retires, what officially replaces what, where the pricing and behavior profiles diverge, a per-workload decision tree for choosing a target model, and a concrete audit checklist to run before July 23. Every date and model ID below comes from OpenAI’s own developer documentation.

OpenAI’s API deprecations page lists the wave under a single April 22, 2026 notice titled “Legacy GPT model snapshots,” stating that access to a set of older models will be shut down on the listed dates “to improve reliability and make it easier for developers to choose the right models.” The July 23 batch spans several unrelated product families — coding agents, chat aliases, deep research, computer use, and legacy audio and search-preview snapshots — batched onto one shared shutdown date.

The Codex-lineage models are the headline for engineering teams. All five pre-5.3 Codex variants retire together: gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini, and gpt-5.2-codex. Alongside them go gpt-5-chat-latest and gpt-5.1-chat-latest, the two deep-research variants (o3-deep-research-2025-06-26 and o4-mini-deep-research-2025-06-26), and computer-use-preview-2025-03-11 — plus a handful of near-duplicate legacy audio, realtime, and search-preview aliases that round out the table.

One structural detail worth noticing: the batch mixes notice tiers. computer-use-preview-2025-03-11 is a preview-named model, which OpenAI’s policy says can retire with “much shorter notice, such as 2 weeks” — yet it landed on the same July 23 date as the generally-available Codex snapshots. OpenAI appears to batch unrelated notice tiers onto shared shutdown dates for operational simplicity, which is good news for anyone maintaining a migration calendar: fewer distinct deadlines to track.

The deprecations page doubles as OpenAI’s published policy on how much warning developers get. Generally-available models receive at least six months. Specialized variants of generally-available models — the policy explicitly names Codex variants such as gpt-5.3-codex, chat variants such as gpt-5.1-chat-latest, and deep-research variants such as o3-deep-research as the example category — receive at least three months. Preview models can go with as little as two weeks.

Measured against that policy, this wave is exactly on-spec: April 22 to July 23 is 92 days, a hair over the three-month minimum for the specialized tier that dominates the list. That’s the framing worth internalizing — this is routine platform hygiene executed to a published schedule, and the schedule itself tells you how long future Codex-variant pins can be expected to live.

Here’s the trap that will send someone on your team chasing the wrong deadline. OpenAI is currently running two different deprecation mechanisms that both touch Codex models, and they mean different things.

The first is the API-level shutdown described above: on July 23, API calls to the retired snapshot IDs stop working. Hard cutoff, listed date, listed replacement.

The second is a product-level notice in OpenAI’s Codex models documentation, which states: “The gpt-5.2 and gpt-5.3-codex models are deprecated in Codex when you sign in with ChatGPT. If your scripts, configuration files, or codex exec --model commands still reference deprecated models, update them to the latest model listed above.” That’s a soft steer inside the Codex app sign-in path — the docs point users toward the current GPT-5.6 family and GPT-5.5/5.4 — not an API shutdown with a date attached.

The distinction matters most for teams that touch Codex through multiple surfaces. The same config.toml model entry is shared across the ChatGPT desktop app, the Codex CLI, and the IDE extension — so one config decision propagates everywhere. If you’ve customized model pins there, our deep dive on Codex CLI’s config.toml and sandbox profiles covers exactly where those entries live and how profiles override each other.

OpenAI’s deprecations page is one long, flat table mixing Codex, audio, realtime, search-preview, and deep-research rows in announcement order. The table below reorganizes the July 23 wave by workload family, with OpenAI’s official replacement, the replacement’s published pricing, and the alternative target worth evaluating if the official mapping doesn’t fit your workload. Pricing is per 1M tokens, input / output, standard tier.

Two things in that table deserve a second look. First, the deep-research mapping is a price-tier change: gpt-5.5-pro runs $30 / $180 per 1M tokens — so long-running research workloads need re-modeling, not just repointing. Second, gpt-5.5 carries a long-context surcharge: prompts over 272K input tokens are billed at 2x input / 1.5x output for the full session. A Codex workload that routinely stuffs large repos into context can land on a different effective price than the headline rate suggests. Our GPT-5.2-to-5.5 migration playbook covers the reasoning-effort and behavior deltas in detail — the same gotchas apply when you’re arriving from a Codex snapshot.

The official replacement column is a default, not a decision. Depending on what your retiring snapshot actually does all day, there are four defensible targets — including one option OpenAI’s table doesn’t surface at all, and one that only became available five days ago with GPT-5.6’s July 9 GA.

The GPT-5.3-Codex branch deserves the caveat spelled out: it survives the July 23 API shutdown — it isn’t on the list — but the Codex app already steers ChatGPT-signed-in users away from it. Reading OpenAI’s own policy language, Codex variants sit in the three-months-notice tier, so a team adopting it today should assume a shorter remaining life than a GA general-purpose model and keep the exit path cheap. If your workloads lean on Codex’s subagents and long autonomous runs, that’s exactly the profile worth re-benchmarking on both GPT-5.3-Codex and the general-purpose line before deciding.

The failure mode here is mundane: a pinned model string in a file nobody has opened since January. The fix is equally mundane — grep for the five literal Codex IDs (plus the chat-latest and deep-research IDs if you use them) across everything that can hold configuration. Concretely:

• config.toml model entries. One shared config drives the ChatGPT desktop app, Codex CLI, and IDE extension — check the model key and any per-profile overrides. The Codex CLI’s Rust-based config changes moved some of these locations, so don’t trust memory of where they used to live.
• codex exec --model invocations in CI pipelines, cron jobs, git hooks, and shell aliases — the places most likely to carry a stale pin silently.
• API integrations and agent backends that pass a model ID explicitly — SDK calls, gateway configs, router rules, and environment variables like OPENAI_MODEL.
• Fine-tuned models. Check whether any fine-tunes in your org are tied to bases on a deprecation list — the October 23 wave explicitly includes fine-tuned variants of gpt-3.5-turbo, gpt-4, and gpt-4.1-nano, and trade coverage flags fine-tunes on deprecated bases as losing access with their base model.

Trade coverage puts the stakes bluntly — as TheRouter.ai’s analysis of the wave framed it: “If you’re running any agent that calls gpt-5-codex or gpt-5.1-codex as a backend, that workflow breaks on July 23.” That’s trade-press commentary rather than OpenAI’s wording, but it’s an accurate reading of what a hard API shutdown means for pinned agent backends.

The same April 22 notice already schedules a second, larger wave: October 23, 2026 retires 12+ more legacy models — including gpt-3.5-turbo-0125, gpt-4-0613, gpt-4-turbo, gpt-4o-2024-05-13, gpt-image-1, o1 and o1-pro snapshots, and o3-mini/o4-mini snapshots — plus fine-tuned variants of gpt-3.5-turbo, gpt-4, and gpt-4.1-nano. A separate June 11 notice then schedules the original gpt-5 family snapshots and o3/o3-pro for shutdown on December 11, 2026.

And the model table isn’t the whole picture. A May 8 notice separately retired gpt-5.2-chat-latest and gpt-5.3-chat-latest on a later summer window, and June 3 notices put whole platform surfaces on the clock: Agent Builder shuts down November 30, 2026, and the Evals platform’s dashboard and API follow the same date, with existing evals read-only from October 31. Taken together, mid-2026 OpenAI is running a coordinated platform-simplification program — consolidating model variants, aliases, and side products onto a smaller, current surface.

The trend worth reading here isn’t “OpenAI kills models” — it’s the consolidation direction. Every specialized Codex snapshot in this wave maps to a general-purpose model, not a newer Codex snapshot. Combined with the policy language that gives specialized variants half the notice of GA models, the signal is that variant-specific model IDs are becoming short-lived implementation details, while the durable interface is the current general-purpose family plus a config layer that can repoint quickly. Given the scale of Codex’s user base, even a small percentage of stale pins translates into a lot of broken CI jobs on July 24.

Looking forward, the practical posture is to treat model migrations as a repeatable process rather than an event. Teams that maintain a standing eval suite — real tasks from their own repos, scored automatically — can qualify a replacement model in days and ride every deprecation wave cheaply; teams that don’t will re-litigate “is the new model good enough?” from scratch every quarter. We’ve written up how to build an LLM eval harness that qualifies new models for exactly this recurring situation. With October 23 and December 11 already on the calendar, that capability pays for itself twice more this year — and if you’d rather not build the qualification pipeline alone, our AI transformation engagements start with exactly this kind of model-portfolio and migration audit.