On this page · 11 sections
- What OpenAI removes on 1 April 2027
- The named replacement is already one version behind
- Price and specification comparison
- Behaviour differences that break code
- Worked cost examples
- Which replacement to pick
- Migration test plan
- Codex CLI, CI and gateway changes
- Azure and EU data residency
- How eCorpIT can help
- References
Summary. OpenAI's deprecations page carries a notice dated 1 October 2026: gpt-5.3-codex, gpt-5.1 (snapshot gpt-5.1-2025-11-13) and gpt-5.4-nano (snapshot gpt-5.4-nano-2026-03-17) leave the API on 1 April 2027, after six months' notice. OpenAI names gpt-6-sol as the replacement for GPT-5.3-Codex and GPT-5.1, and gpt-6-luna for GPT-5.4 nano. Prices move both ways. GPT-5.3-Codex lists at $1.75 input and $14 output per million tokens; GPT-6 Sol costs $2 and $10, so output is 29% cheaper, input 14% dearer, and cache writes now bill at $2.50. GPT-5.1 at $1.25 and $10 becomes $2 and $10. GPT-5.4 nano at $0.20 and $1.25 becomes $0.10 and $0.50 on Luna. The behaviour changes matter more: GPT-6 Sol and Luna default to medium reasoning where GPT-5.1 and GPT-5.4 nano default to none, allow Chat Completions function calling only at none, and bill prompts above 272,000 input tokens at 2x input and 1.5x output. OpenAI's GPT-6 Sol page already points to GPT-6.1 Sol as the newer Sol model, at the same $2 and $10 but without none effort. Pick a target per workload, test it against a golden set and finish the rollout by mid-March 2027.
About this article. This guide is for engineering leads and developers who call these three models through the OpenAI API, Codex or Azure. On 7 November 2026 we checked every model ID, date, price and parameter against OpenAI's deprecations, pricing, model, changelog, guide and Codex pages and Microsoft's Foundry retirement schedule. Vendors change prices often, so re-check before you commit. Not legal or financial advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.
What OpenAI removes on 1 April 2027
The notice is the 1 October 2026 entry on OpenAI's deprecations page, and the three model pages now carry a Deprecated label.
| Model or alias | Snapshot | API shutdown | OpenAI's recommended replacement |
|---|---|---|---|
| gpt-5.3-codex | gpt-5.3-codex (single snapshot) | 1 April 2027 | gpt-6-sol |
| gpt-5.1 | gpt-5.1-2025-11-13 | 1 April 2027 | gpt-6-sol |
| gpt-5.4-nano | gpt-5.4-nano-2026-03-17 | 1 April 2027 | gpt-6-luna |
- Six months is double the minimum for one model and the floor for two. OpenAI promises at least six months for generally available models and at least three for specialised variants, naming
gpt-5.3-codexas its example of a Codex variant. GPT-5.3-Codex got twice its minimum; GPT-5.1 and GPT-5.4 nano got exactly the 182 days the GA rule requires. Our comparison of AI model retirement policies sets this against other vendors.
- Aliases and snapshots go together. The notice lists both forms for GPT-5.1 and GPT-5.4 nano, so both stop working on the same day.
- You can watch it in code. The List models endpoint (
GET /v1/models) and the Retrieve model endpoint return an optionalshutdown_datefield as aYYYY-MM-DDstring, or null when no date is set. Check it nightly and fail the build when a pinned model has a date inside 90 days.
OpenAI also says some developers may be able to arrange dedicated capacity after shutdown through its sales team. Treat that as an exception for large accounts, not a plan. GPT-4o and GPT-4.1 go first, on 14 January 2027, also to Luna, as our GPT-4o and GPT-4.1 shutdown guide explains.
The Codex-branded line ends here on the API
GPT-5.3-Codex is the only Codex model in OpenAI's API catalogue that has not already shut down. The last five, from gpt-5-codex to gpt-5.2-codex, went on 23 July 2026 with gpt-5.6-sol or gpt-5.6-terra as substitutes. The pricing page lists gpt-5.3-codex as the only model in its Codex category, and OpenAI's Codex documentation now recommends GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna, all general models.
In Codex itself, the date depends on sign-in. OpenAI's Codex models page says gpt-5.3-codex is already deprecated in Codex for ChatGPT sign-in, while API-key sessions follow the model availability of your API organisation and project. Teams running codex exec in CI with an API key therefore meet the 1 April date.
The named replacement is already one version behind
The GPT-6 Sol model page sends readers to GPT-6.1 Sol as the newer Sol model; OpenAI released the two a week apart, on 22 and 29 September 2026. GPT-6.1 Sol keeps the $2 input and $10 output price, halves cached input to $0.10, and is positioned by OpenAI as close to Astra-level performance at a lower cost. On 7 November 2026 the deprecations page lists no shutdown date for gpt-6-sol, and the GA policy promises at least six months' notice for any future one.
The comparison draws on both model pages and the GPT-6 guide.
| Point | GPT-6 Sol | GPT-6.1 Sol |
|---|---|---|
| Reasoning effort values | none, low, medium (default), high, xhigh, max | low, medium (default), high, xhigh, max |
| Tool calling | Responses API; Chat Completions only at none | Responses API only |
| Cached input per 1M tokens | $0.20 | $0.10 |
| Knowledge cutoff | 20 April 2026 | 30 April 2026 |
| Fast mode with EU data residency | Not available | Available |
| Ultrafast mode | Not offered | 6x Standard |
For coding agents moving off GPT-5.3-Codex, GPT-6.1 Sol is the natural target: GPT-5.3-Codex already runs only at low to xhigh effort and only on the Responses API, so its effort settings and API calls carry over, and cache reads get cheaper. For GPT-5.1 code that calls functions through Chat Completions at none effort, only GPT-6 Sol and GPT-6 Luna keep that path open. GPT-6.1 Terra supports none but, like GPT-6.1 Sol, needs the Responses API for any tool call.
Price and specification comparison
Prices are OpenAI Standard tier, per million tokens, for prompts up to 272,000 input tokens, from the model pages and the pricing page; cache writes are for the default 30-minute retention.
| Model | Input | Cached input | Cache write | Output | Context window | Max output | Knowledge cutoff | Batch API |
|---|---|---|---|---|---|---|---|---|
| gpt-5.3-codex (retiring) | $1.75 | $0.175 | None listed | $14 | 400,000 | 128,000 | 31 Aug 2025 | No |
| gpt-5.1 (retiring) | $1.25 | $0.125 | None listed | $10 | 400,000 | 128,000 | 30 Sep 2024 | Yes |
| gpt-5.4-nano (retiring) | $0.20 | $0.02 | None listed | $1.25 | 400,000 | 128,000 | 31 Aug 2025 | Yes |
| gpt-6-sol | $2 | $0.20 | $2.50 | $10 | 1,050,000 | 128,000 | 20 Apr 2026 | Yes |
| gpt-6.1-sol | $2 | $0.10 | $2.50 | $10 | 1,050,000 | 128,000 | 30 Apr 2026 | Yes |
| gpt-6-luna | $0.10 | $0.01 | $0.125 | $0.50 | 1,050,000 | 128,000 | 18 May 2026 | Yes |
Rules that change the bill on the GPT-6 models:
- Long-prompt surcharge. Prompts above 272,000 input tokens pay 2x input and cache rates and 1.5x output for the whole request. The retiring models have 400,000-token windows, so an agent that starts filling the new 1,050,000-token window can double its input cost without any error. In Codex, GPT-6.1 Sol defaults to a 500,000-token window, and OpenAI exempts subscription usage from the surcharge, which implies API-key runs in CI pay it.
- Cache writes. Writes bill at 1.25x the uncached input rate for 30-minute retention and 2.5x for 12-hour retention. These are total rates, not add-ons, and explicit-only caching with no breakpoints creates no writes. The retiring models' pages show no cache-write price.
- Tiers. Batch and Flex are 50% of Standard, Fast mode is 2x and GPT-6.1 Sol's Ultrafast mode is 6x. GPT-5.3-Codex offered Fast at $3.50 and $28 but no Batch, so nightly review jobs can move to Batch at $1 and $5 on either Sol.
- Regional processing. Data residency endpoints add 10% for models released on or after 5 March 2026, which covers the three GPT-6 targets and GPT-5.4 nano.
GPT-6.1 Terra, at $1 input, $0.05 cached and $5 output, sits between Sol and Luna and supports none effort. It is not a named replacement, but it belongs on the shortlist for GPT-5.1 work that does not need Sol.
Behaviour differences that break code
Most failures in this migration will come from parameters, not model quality. OpenAI's GPT-6 guide lists the changes; the table maps them to the retiring models using each model page.
| Area | Retiring model behaviour | GPT-6 behaviour | What to change |
|---|---|---|---|
| Default reasoning | gpt-5.1 and gpt-5.4-nano default to none; gpt-5.3-codex supports low to xhigh | GPT-6 Sol, GPT-6.1 Sol and Luna default to medium | Set effort on every call; an unset effort becomes a silent cost and latency rise |
| none effort | Supported on gpt-5.1 and gpt-5.4-nano | Supported on GPT-6 Sol, Luna and GPT-6.1 Terra; not on GPT-6.1 Sol | Map none to low if you pick GPT-6.1 Sol |
| Tool calls in Chat Completions | gpt-5.1 and gpt-5.4-nano support them; gpt-5.3-codex has no Chat Completions | GPT-6 Sol and Luna only at none; GPT-6.1 Sol and Terra need Responses | Move tool-using code to the Responses API |
| Sampling parameters | Code written for none effort may send them | Unsupported whenever effort is not none | Drop temperature, top_p and top_logprobs, plus logprobs in Chat Completions |
| Prompt caching | prompt_cache_retention | prompt_cache_options.ttl, default 30m | Rename the parameter and budget for cache writes |
| Responses tools | gpt-5.4-nano lacks computer use and tool search; the gpt-5.3-codex page no longer lists its tools | All ten built-in tools listed, including MCP, computer use and tool search | Re-test tool selection; more tools means more ways to pick the wrong one |
Two more points deserve a test case each. OpenAI's prompting advice, written around GPT-6 Astra but offered as a starting point for the whole family, describes a model more likely to stop and ask a clarifying question and more sensitive to instructions in skills and files such as AGENTS.md. An unattended coding agent that pauses for approval is a failed CI job. And the knowledge cutoff moves forward about eight months from GPT-5.3-Codex and about 19 months from GPT-5.1, which can change library and API suggestions in generated code.
Worked cost examples
These use list prices only, keep token counts constant and ignore reasoning-token differences. eCorpIT did not benchmark these models; measure your own traffic.
Scenario A: coding agent on gpt-5.3-codex. 2,000 million input tokens a month, 80% from cache (1,600 million), 400 million uncached, and 60 million output tokens. For the GPT-6 models the uncached line runs from no cache writes (400 x $2 = $800) to every uncached token written to a 30-minute cache (400 x $2.50 = $1,000).
| Line item | gpt-5.3-codex | gpt-6-sol | gpt-6.1-sol |
|---|---|---|---|
| Cache reads, 1,600M | 1,600 x $0.175 = $280 | 1,600 x $0.20 = $320 | 1,600 x $0.10 = $160 |
| Uncached input, 400M | 400 x $1.75 = $700 | $800 to $1,000 | $800 to $1,000 |
| Output, 60M | 60 x $14 = $840 | 60 x $10 = $600 | 60 x $10 = $600 |
| Monthly total | $1,820 | $1,720 to $1,920 | $1,560 to $1,760 |
On list prices the move is roughly cost-neutral, and GPT-6.1 Sol comes out cheapest because of its $0.10 cache reads. The real variable is effort: if the agent runs GPT-6.1 Sol at xhigh and output doubles to 120 million tokens, the total rises by 60 x $10 = $600.
Scenario B: gpt-5.1 at none effort for summarisation. 1,000 million input and 100 million output tokens, no caching.
| Target | Arithmetic | Monthly total | Change |
|---|---|---|---|
| gpt-5.1 (today) | 1,000 x $1.25 + 100 x $10 | $2,250 | Baseline |
| gpt-6-sol | 1,000 x $2 + 100 x $10 | $3,000 | 33% more |
| gpt-6.1-terra | 1,000 x $1 + 100 x $5 | $1,500 | 33% less |
| gpt-6-luna | 1,000 x $0.10 + 100 x $0.50 | $150 | 93% less |
Following OpenAI's mapping costs a third more here. If a GPT-5.1 workload does not need Sol-level reasoning, test Terra and Luna first, at none effort.
Scenario C: gpt-5.4-nano for extraction. 5,000 million input and 500 million output tokens. Today: 5,000 x $0.20 + 500 x $1.25 = $1,625. On GPT-6 Luna: 5,000 x $0.10 + 500 x $0.50 = $750, a 54% saving, provided you set effort to none so Luna does not reason at its default medium.
Which replacement to pick
| Pick | If |
|---|---|
| gpt-6.1-sol | You run coding agents, PR review or CI repair on gpt-5.3-codex, already use the Responses API and can live without none effort |
| gpt-6-sol | You need OpenAI's named replacement for audit reasons, or GPT-5.1 code uses Chat Completions function calling at none |
| gpt-6.1-terra | Your GPT-5.1 workload is writing, extraction or screen understanding, its tool calls can move to Responses, and cost matters more than peak reasoning |
| gpt-6-luna | You run gpt-5.4-nano for classification, extraction, ranking or sub-agents, or GPT-5.1 at none on simple tasks |
| Azure OpenAI as a bridge | You already use Azure and need more time; see the dates below |
For a cross-vendor view, our guide to the best LLM for coding in 2027 compares GPT-6.1 Sol with Claude Opus 5.5 and Gemini 4 Argon.
Migration test plan
From 7 November 2026 there are 145 days to 1 April 2027, enough for a measured test rather than a model-string swap.
- Inventory every reference. Search code, infrastructure config, prompt files,
config.toml, CI workflows and gateway routes for all five IDs above. Catchgpt-5,gpt-5-mini,gpt-5-nanoando3in the same pass: OpenAI shuts those down on 11 December 2026. Cross-check against your usage dashboard.
- Add the shutdown check. Call
GET /v1/modelsnightly and alert on any pinned model with ashutdown_date.
- Build a golden set per workload. For coding agents, take 50 to 200 real tasks with tests and track pass rate, diff size, tool-call errors, tokens per task and wall time. For GPT-5.1 and nano, label 100 to 300 production inputs with expected outputs or schema checks.
- Record the baseline now, on the retiring model at your production effort level, while it still exists.
- Fix parameters before judging quality. Apply the behaviour table: explicit effort, sampling parameters removed, caching parameter renamed, tool calls on Responses.
- Run two or three candidates at two effort levels, for example GPT-6.1 Sol at
mediumandhighand GPT-6 Sol atmedium. Set pass gates first: task success within an agreed margin of baseline, no rise in schema failures, and cost per solved task and p95 latency within budget.
- Shadow, then canary. Mirror 5% of traffic without acting on the output, then route 5%, 25% and 100%. Keep a second live model as rollback, because the old one will not exist after 1 April.
- Watch the agent's manners. Count clarifying questions, approval pauses and early stops, and use OpenAI's initiative and follow-through prompts if the count rises.
- Monitor after cut-over. Alert on any error that names a retired model, and on token-per-task drift.
| Window | Milestone |
|---|---|
| By 30 November 2026 | Inventory complete, shutdown check live |
| 1 to 31 December 2026 | Golden sets built, baselines recorded on retiring models |
| 4 to 29 January 2027 | Parameter fixes, candidate runs, target chosen per workload |
| 1 to 26 February 2027 | Shadow traffic, then 5% and 25% canaries |
| By 12 March 2027 | 100% of traffic on new models |
| 1 April 2027 | Shutdown; retired-model error alerts on |
Codex CLI, CI and gateway changes
Codex takes its model from config.toml, the --model flag or its -m alias, or a workspace default; the desktop app, CLI and IDE extension share one file. OpenAI's examples use this form:
# ~/.codex/config.toml, or a profile file such as ~/.codex/ci.config.toml
model = "gpt-6.1-sol"
codex exec -m gpt-6.1-sol "Review the current changes"
Passing --profile layers a profile file over the base config, which keeps a CI model separate from developer defaults. The Codex GitHub Action (openai/codex-action@v1) takes model, effort and codex-version inputs; empty values fall back to defaults and the latest CLI. Set the model explicitly and pin codex-version so a CLI upgrade and a model change do not land in the same week. Check builds too: GPT-6.1 Terra requires Codex CLI 0.159.0 or later.
Gateways must expose a Responses-compatible endpoint, because OpenAI says current Codex releases do not support a wire_api value of chat or a Chat Completions-only endpoint. An AI gateway such as LiteLLM, Portkey, Kong or TrueFoundry turns the cut-over into a routing change and shows per-model spend in one place.
Azure and EU data residency
Azure dates are later. Microsoft Foundry's model retirement schedule, updated on 23 September 2026, retires gpt-5.1 (2025-11-13) on 13 July 2027, gpt-5.3-codex (2026-02-24) on 24 August 2027 and gpt-5.4-nano (2026-03-17) on 21 September 2027, with no replacement named yet for any of them. Standard deployments auto-upgrade according to their versionUpgradeOption setting, provisioned deployments never do, and the Models API exposes lifecycleStatus and per-SKU deprecationDate fields. Our Azure OpenAI Codex retirements guide covers the rest.
EU data residency. GPT-6 Sol supports EU data residency with Standard, Flex and Batch processing but not with Fast mode, while GPT-6.1 Sol and GPT-6 Luna support it with Fast as well. A team that needs Fast mode inside the EU should pick GPT-6.1 Sol.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our LLM migration and cost-optimisation service measures cost per task on your real traffic, builds the evaluation and routing layer, and moves workloads off retiring models such as these three. Talk to us at /contact-us/.
Last updated: 7 November 2026.
